{
  "schema_version": "1.0.0",
  "license": "CC-BY-4.0",
  "attribution": "CanIEmbody (https://caniembody.pages.dev)",
  "dataset_version": "1.0.0-seed",
  "generated": "2026-10-25",
  "data": [
    {
      "id": "evo1-jetson-orin-nano-8gb-onnx-split-fp16-1cam-eetmie-2026-09",
      "policy": "evo1",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "EVO1 \"bootstrap\" export: OpenGVLab/InternVL3-1B-hf initializer with a randomly initialized action head (infrastructure-only, not deployable)",
      "date": "2026-09-05",
      "runtime": "onnx-split",
      "runtime_note": "Split ONNX graphs on ONNX Runtime (onnxruntime-gpu 1.24) with the TensorRT execution provider (FP16); superseded in the same repo by the pure-TensorRT runs.",
      "precision": "fp16",
      "cameras": 1,
      "inference_latency_ms": {
        "p50": 289.18,
        "p95": 292.12
      },
      "inference_hz": 3.45,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Super 8 GB developer kit",
        "jetpack": "7.2 (L4T R39.2.x, per the repo host-setup doc at this commit)",
        "power_mode": "MAXN_SUPER, pinned clocks",
        "cuda": "13.2",
        "os": "Ubuntu 24.04",
        "onnxruntime-gpu": "1.24",
        "tensorrt": "10.16.2 (system, via ONNX Runtime TensorRT EP)"
      },
      "note": "Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Untrained action head: measures inference cost of the EVO1 architecture only; the author marks it as never to control a robot. Author labels the precision \"mixed FP16\".",
      "source_ref": "ev-eetmie-readme-2026-09",
      "evidence": [
        {
          "id": "ev-eetmie-readme-2026-09",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/README.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README \"Measured fit\" table at commit b6185e1 (PyTorch FP32 vs split ONNX FP16 on Orin Nano 8 GB Super).",
          "quote": "| EVO1 bootstrap split ONNX mixed FP16 | 1 | 289.18 ms | 292.12 ms | 3.45 Hz |"
        },
        {
          "id": "ev-eetmie-results-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/RESULTS.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Generated results at the same commit: speed, footprint and control-rate fit per run."
        },
        {
          "id": "ev-eetmie-host-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/01-host-setup.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Board state at that commit: JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.",
          "quote": "JetPack 7.2 is L4T R39.2.0: Ubuntu 24.04, kernel 6.8, Python 3.12, CUDA 13.2, cuDNN 9.20,"
        }
      ]
    },
    {
      "id": "evo1-jetson-orin-nano-8gb-onnx-split-fp16-2cam-eetmie-2026-09",
      "policy": "evo1",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "zuoxingdong/evo1_libero",
      "date": "2026-09-05",
      "runtime": "onnx-split",
      "runtime_note": "Split ONNX graphs on ONNX Runtime (onnxruntime-gpu 1.24) with the TensorRT execution provider (FP16); superseded in the same repo by the pure-TensorRT runs.",
      "precision": "fp16",
      "cameras": 2,
      "inference_latency_ms": {
        "p50": 414.67,
        "p95": 424.72
      },
      "inference_hz": 2.41,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Super 8 GB developer kit",
        "jetpack": "7.2 (L4T R39.2.x, per the repo host-setup doc at this commit)",
        "power_mode": "MAXN_SUPER, pinned clocks",
        "cuda": "13.2",
        "os": "Ubuntu 24.04",
        "onnxruntime-gpu": "1.24",
        "tensorrt": "10.16.2 (system, via ONNX Runtime TensorRT EP)"
      },
      "note": "Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Author labels the precision \"mixed FP16\". System RAM 6001 MB in use.",
      "source_ref": "ev-eetmie-readme-2026-09",
      "evidence": [
        {
          "id": "ev-eetmie-readme-2026-09",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/README.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README \"Measured fit\" table at commit b6185e1 (PyTorch FP32 vs split ONNX FP16 on Orin Nano 8 GB Super).",
          "quote": "| EVO1 LIBERO split ONNX mixed FP16 | 2 | 414.67 ms | 424.72 ms | 2.41 Hz |"
        },
        {
          "id": "ev-eetmie-results-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/RESULTS.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Generated results at the same commit: speed, footprint and control-rate fit per run."
        },
        {
          "id": "ev-eetmie-host-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/01-host-setup.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Board state at that commit: JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.",
          "quote": "JetPack 7.2 is L4T R39.2.0: Ubuntu 24.04, kernel 6.8, Python 3.12, CUDA 13.2, cuDNN 9.20,"
        }
      ]
    },
    {
      "id": "evo1-jetson-orin-nano-8gb-tensorrt-fp16-2cam-eetmie",
      "policy": "evo1",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "zuoxingdong/evo1_libero",
      "date": "2026-10-13",
      "runtime": "tensorrt",
      "runtime_note": "Split ONNX export, each graph built into its own TensorRT 10.16.2 engine on the board, engines chained and replayed as one CUDA graph, plus ahead-of-time-compiled Triton kernels (eetmie/jetson-orin-nano-vla). Precision as recorded: mixed fp16 (strongly typed); FP32 accumulation by default.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "640x480",
      "inference_latency_ms": {
        "p50": 332.77,
        "p95": 333.63,
        "mean": 333.01
      },
      "inference_hz": 3,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB)",
        "jetpack": "7.2.1",
        "l4t": "R39.2.1",
        "power_mode": "MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz)",
        "tensorrt": "10.16.2.10",
        "os": "Ubuntu 24.04.4 LTS",
        "python": "3.12.3",
        "desktop": "running"
      },
      "note": "Inference cost only; not task success. 300 s run on synthetic 640x480 observations resized by the exported policy to 448x448; 32 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 2.76 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing.",
      "source_ref": "ev-eetmie-readme",
      "evidence": [
        {
          "id": "ev-eetmie-readme",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README results table and benchmark conditions (Orin Nano 8 GB Super, TensorRT).",
          "quote": "| [EVO1 775M LIBERO](results/evo1-libero.json) | [`zuoxingdong/evo1_libero`](https://huggingface.co/zuoxingdong/evo1_libero) | 2 | 332.77 ms | 3.00 Hz | 2.76 GB |"
        },
        {
          "id": "ev-eetmie-conditions",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions stated under the table.",
          "quote": "JetPack 7.2.1, MAXN_SUPER with pinned clocks, desktop running (though headless recommended!). 300-second runs, synthetic observations with default denoising steps and views. RAM is whole-board usage during inference."
        },
        {
          "id": "ev-eetmie-results-json",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/results/evo1-libero.json",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Raw run JSON: latency distribution (p50/p95/mean), board state (L4T R39.2.1, MAXN_SUPER, pinned clocks), TensorRT 10.16.2.10, chunk size, steps."
        }
      ]
    },
    {
      "id": "groot-n1-7-h100-80gb-pytorch-eager-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "h100-80gb",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "pytorch-eager",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 101.5
      },
      "inference_hz": 9.85,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| H100 80GB HBM3 | 101.5 | 53.4 | 22.9 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-h100-80gb-tensorrt-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "h100-80gb",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "tensorrt",
      "runtime_note": "TensorRT full pipeline (backbone and action head engines), baseline execution profile.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 22.9
      },
      "inference_hz": 43.67,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| H100 80GB HBM3 | 101.5 | 53.4 | 22.9 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-h100-80gb-torch-compile-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "h100-80gb",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "torch-compile",
      "runtime_note": "torch.compile max-autotune on the action head; backbone stays eager.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 53.4
      },
      "inference_hz": 18.73,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| H100 80GB HBM3 | 101.5 | 53.4 | 22.9 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-jetson-agx-thor-pytorch-eager-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "jetson-agx-thor",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "pytorch-eager",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 106.8
      },
      "inference_hz": 9.36,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op",
        "power_mode": "MAXN with jetson_clocks",
        "jetpack": "7.2",
        "cuda": "13.2",
        "torch": "2.13.0"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| AGX Thor (MAXN) | 106.8 | 100.5 | 76.8 | 35.6 |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        },
        {
          "id": "ev-groot-jetson-stack",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Software stack used on Thor.",
          "quote": "Thor and Orin use JetPack 7.2 / CUDA 13.2 with PyTorch 2.13.0"
        }
      ]
    },
    {
      "id": "groot-n1-7-jetson-agx-thor-tensorrt-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "jetson-agx-thor",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "tensorrt",
      "runtime_note": "TensorRT full pipeline (backbone and action head engines), baseline execution profile.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 76.8
      },
      "inference_hz": 13.02,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op",
        "power_mode": "MAXN with jetson_clocks",
        "jetpack": "7.2",
        "cuda": "13.2",
        "torch": "2.13.0"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| AGX Thor (MAXN) | 106.8 | 100.5 | 76.8 | 35.6 |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        },
        {
          "id": "ev-groot-jetson-stack",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Software stack used on Thor.",
          "quote": "Thor and Orin use JetPack 7.2 / CUDA 13.2 with PyTorch 2.13.0"
        }
      ]
    },
    {
      "id": "groot-n1-7-jetson-agx-thor-tensorrt-mixed-2cam-isaac-groot-nvfp4",
      "policy": "groot-n1-7",
      "compute": "jetson-agx-thor",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "tensorrt",
      "runtime_note": "TensorRT full pipeline with --execution-profile optimized (cross-attention K/V caching, projection fusion, CUDA graphs) and --quantization mixed_nvfp4 (Jetson AI Lab recipe).",
      "precision": "mixed",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 35.6
      },
      "inference_hz": 28.09,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op",
        "power_mode": "MAXN with jetson_clocks",
        "jetpack": "7.2",
        "cuda": "13.2",
        "torch": "2.13.0"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Mixed NVFP4 quantization (enum has no fp4 value).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| AGX Thor (MAXN) | 106.8 | 100.5 | 76.8 | 35.6 |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        },
        {
          "id": "ev-groot-jetson-stack",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Software stack used on Thor.",
          "quote": "Thor and Orin use JetPack 7.2 / CUDA 13.2 with PyTorch 2.13.0"
        }
      ]
    },
    {
      "id": "groot-n1-7-jetson-agx-thor-torch-compile-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "jetson-agx-thor",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "torch-compile",
      "runtime_note": "torch.compile max-autotune on the action head; backbone stays eager.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 100.5
      },
      "inference_hz": 9.95,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op",
        "power_mode": "MAXN with jetson_clocks",
        "jetpack": "7.2",
        "cuda": "13.2",
        "torch": "2.13.0"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| AGX Thor (MAXN) | 106.8 | 100.5 | 76.8 | 35.6 |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        },
        {
          "id": "ev-groot-jetson-stack",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Software stack used on Thor.",
          "quote": "Thor and Orin use JetPack 7.2 / CUDA 13.2 with PyTorch 2.13.0"
        }
      ]
    },
    {
      "id": "groot-n1-7-jetson-orin-nano-8gb-tensorrt-fp16-3cam-eetmie",
      "policy": "groot-n1-7",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "nvidia/GR00T-N1.7-3B",
      "date": "2026-10-13",
      "runtime": "tensorrt",
      "runtime_note": "Split ONNX export, each graph built into its own TensorRT 10.16.2 engine on the board, engines chained and replayed as one CUDA graph, plus ahead-of-time-compiled Triton kernels (eetmie/jetson-orin-nano-vla). Precision as recorded: mixed fp16 (strongly typed); FP32 accumulation by default.",
      "precision": "fp16",
      "cameras": 3,
      "resolution": "640x480",
      "inference_latency_ms": {
        "p50": 273.7,
        "p95": 275.4,
        "mean": 273.58
      },
      "inference_hz": 3.66,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB)",
        "jetpack": "7.2.1",
        "l4t": "R39.2.1",
        "power_mode": "MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz)",
        "tensorrt": "10.16.2.10",
        "os": "Ubuntu 24.04.4 LTS",
        "python": "3.12.3",
        "desktop": "running"
      },
      "note": "Inference cost only; not task success. 300 s run on synthetic 640x480 observations resized by the exported policy to 256x256; 4 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 5.78 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing. Three camera views, each with two frames (GR00T N1.7 temporal input).",
      "source_ref": "ev-eetmie-readme",
      "evidence": [
        {
          "id": "ev-eetmie-readme",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README results table and benchmark conditions (Orin Nano 8 GB Super, TensorRT).",
          "quote": "| [GR00T N1.7 3B](results/groot-n17-base.json) | [`nvidia/GR00T-N1.7-3B`](https://huggingface.co/nvidia/GR00T-N1.7-3B) | 3 (×2 frames) | 273.70 ms | 3.66 Hz | 5.78 GB |"
        },
        {
          "id": "ev-eetmie-conditions",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions stated under the table.",
          "quote": "JetPack 7.2.1, MAXN_SUPER with pinned clocks, desktop running (though headless recommended!). 300-second runs, synthetic observations with default denoising steps and views. RAM is whole-board usage during inference."
        },
        {
          "id": "ev-eetmie-results-json",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/results/groot-n17-base.json",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Raw run JSON: latency distribution (p50/p95/mean), board state (L4T R39.2.1, MAXN_SUPER, pinned clocks), TensorRT 10.16.2.10, chunk size, steps."
        }
      ]
    },
    {
      "id": "groot-n1-7-l40-pytorch-eager-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "l40",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "pytorch-eager",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 123.7
      },
      "inference_hz": 8.08,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| L40 | 123.7 | 66.8 | 35.0 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-l40-tensorrt-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "l40",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "tensorrt",
      "runtime_note": "TensorRT full pipeline (backbone and action head engines), baseline execution profile.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 35
      },
      "inference_hz": 28.57,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| L40 | 123.7 | 66.8 | 35.0 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-l40-torch-compile-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "l40",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "torch-compile",
      "runtime_note": "torch.compile max-autotune on the action head; backbone stays eager.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 66.8
      },
      "inference_hz": 14.97,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| L40 | 123.7 | 66.8 | 35.0 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-rtx-pro-6000-blackwell-server-pytorch-eager-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "rtx-pro-6000-blackwell-server",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "pytorch-eager",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 75.7
      },
      "inference_hz": 13.21,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| RTX Pro 6000 Blackwell Server Edition (96 GB) | 75.7 | 52.1 | 24.7 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-rtx-pro-6000-blackwell-server-tensorrt-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "rtx-pro-6000-blackwell-server",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "tensorrt",
      "runtime_note": "TensorRT full pipeline (backbone and action head engines), baseline execution profile.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 24.7
      },
      "inference_hz": 40.49,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| RTX Pro 6000 Blackwell Server Edition (96 GB) | 75.7 | 52.1 | 24.7 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "groot-n1-7-rtx-pro-6000-blackwell-server-torch-compile-bf16-2cam-isaac-groot",
      "policy": "groot-n1-7",
      "compute": "rtx-pro-6000-blackwell-server",
      "checkpoint": "nvidia/GR00T-N1.7-LIBERO (libero_10)",
      "date": "2026-10-21",
      "runtime": "torch-compile",
      "runtime_note": "torch.compile max-autotune on the action head; backbone stays eager.",
      "precision": "bf16",
      "cameras": 2,
      "resolution": "256x256",
      "batch_size": 1,
      "inference_latency_ms": {
        "p50": 52.1
      },
      "inference_hz": 19.19,
      "action_chunk": 40,
      "outcome": "runs",
      "platform": {
        "torch_threads": "8 intra-op / 1 inter-op"
      },
      "note": "Inference cost only; not task success. Median of 60 timed calls (3 repetitions x 20 after 10 warmup), fixed inputs, includes preprocessing and action decoding, excludes RPC. Precision is the repo default (unquantized BF16).",
      "source_ref": "ev-groot-deploy-bench",
      "evidence": [
        {
          "id": "ev-groot-deploy-bench",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "NVIDIA Isaac-GR00T deployment README benchmark table (GR00T N1.7 LIBERO, median ms, PyTorch eager / torch.compile / TensorRT full pipeline / TensorRT optimized + mixed NVFP4).",
          "quote": "| RTX Pro 6000 Blackwell Server Edition (96 GB) | 75.7 | 52.1 | 24.7 | — |"
        },
        {
          "id": "ev-groot-deploy-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions for the table.",
          "quote": "Median complete-policy-call latency for the `nvidia/GR00T-N1.7-LIBERO` `libero_10` checkpoint (32 DiT blocks): batch size 1, action horizon 40, four denoising steps, and two 256×256 camera inputs."
        },
        {
          "id": "ev-groot-deploy-bf16-default",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/NVIDIA/Isaac-GR00T/blob/944fe70c4fd54e3fcb80a7d7258a328f9c061176/scripts/deployment/README.md",
          "date": "2026-10-21",
          "verified_on": "2026-10-24",
          "summary": "Unquantized BF16 and baseline execution are the defaults; NVFP4/FP8 are opt-in on Thor.",
          "quote": "Baseline execution and unquantized BF16 remain the defaults."
        }
      ]
    },
    {
      "id": "molmoact2-h100-80gb-pytorch-eager-bf16-2cam-lerobot-docs",
      "policy": "molmoact2",
      "compute": "h100-80gb",
      "checkpoint": "MolmoAct2 (LeRobot port; docs do not name the checkpoint)",
      "date": "2026-08-03",
      "runtime": "pytorch-eager",
      "runtime_note": "LeRobot MolmoAct2 policy, continuous action mode, CUDA graph enabled.",
      "precision": "bf16",
      "cameras": 2,
      "batch_size": 1,
      "action_chunk": 10,
      "vram_gb": 13,
      "outcome": "runs",
      "platform": {
        "os": "Ubuntu 22.04 (repo tested with)",
        "memory_sampling": "nvidia-smi peak",
        "lerobot": "0.6.1"
      },
      "note": "Inference memory only (12.1 GiB peak GPU memory = 13.0 GB); no latency reported. LIBERO with two RGB cameras, chunk_size=10, 8 flow timesteps. Not task success.",
      "source_ref": "ev-lerobot-molmoact2-mem",
      "evidence": [
        {
          "id": "ev-lerobot-molmoact2-mem",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/molmoact2.mdx",
          "date": "2026-08-03",
          "verified_on": "2026-10-24",
          "versions": {
            "lerobot": "0.6.1"
          },
          "summary": "LeRobot v0.6.1 MolmoAct2 docs: peak GPU memory 12.1 GiB for continuous inference with CUDA graphs at batch 1 on one H100 80GB.",
          "quote": "| Inference, continuous, CUDA graph enabled (bs=1) | 12.1 GiB |"
        },
        {
          "id": "ev-lerobot-molmoact2-conditions",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/molmoact2.mdx",
          "date": "2026-08-03",
          "verified_on": "2026-10-24",
          "versions": {
            "lerobot": "0.6.1"
          },
          "summary": "Measurement conditions for the table.",
          "quote": "below were taken on a single NVIDIA H100 80GB with bf16 model loading, LIBERO with two RGB cameras. MolmoAct2 rows use `chunk_size=10`"
        }
      ]
    },
    {
      "id": "pi05-jetson-agx-orin-64gb-llama-cpp-fp16-2cam-eaglevla-pr6",
      "policy": "pi05",
      "compute": "jetson-agx-orin-64gb",
      "checkpoint": "π0.5 LIBERO (openpi pi05_libero) converted to GGUF",
      "date": "2026-10-23",
      "runtime": "llama-cpp",
      "runtime_note": "EagleVLA-Edge: llama.cpp-based VLA engine (GGUF FP16 weights) with CUDA graph reuse, GPU-resident intermediate buffers / KV and unrolled flow-matching denoising.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "224x224",
      "inference_latency_ms": {
        "p50": 271.1
      },
      "inference_hz": 3.69,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin, 'Orin-50W' power mode (AGX Orin 64GB inferred, see ev-orin-board-inferred)",
        "power_mode": "50W (paper label \"Orin-50W\")"
      },
      "note": "Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). Later configuration from the EagleVLA-Edge README (time cache + cuBLASLt GEMM tuning, PR #6); not in the paper. Naive PyTorch π0.5 on the same board per the paper: 1420.8 ms (0.70 Hz) (not filed as a separate run: the baseline precision is not stated).",
      "source_ref": "ev-eaglevla-edge-readme",
      "evidence": [
        {
          "id": "ev-eaglevla-edge-readme",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/PKU-SEC-Lab/EagleVLA-Edge/blob/619e92e878a25df3e9cc0ce87b8f3a61283146e7/README.md",
          "date": "2026-10-23",
          "verified_on": "2026-10-24",
          "summary": "EagleVLA-Edge README P50 latency table for π0.5 on Jetson Orin: 271.1 ms total.",
          "quote": "| + time cache, and GEMM tuning (PR #6) | **66.6** | **137.0** | **67.6** | **271.1** |"
        },
        {
          "id": "ev-orin-board-inferred",
          "tier": "inferred",
          "type": "spec-derivation",
          "url": "https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/SD/PlatformPowerAndPerformance/JetsonOrinNanoSeriesJetsonOrinNxSeriesAndJetsonAgxOrinSeries.html",
          "verified_on": "2026-10-24",
          "summary": "INFERRED board identity: the paper names the device only as 'Jetson Orin' and labels this power setting 'Orin-50W' (Table 1). In NVIDIA's Orin power-mode tables only the AGX Orin 64GB has a 50W mode (AGX Orin 32GB: 15/30/40W; Orin NX/Nano: up to 40W/25W). The paper's own hardware table lists both 'Orin 32GB' and 'Orin 64GB', so this mapping is a deduction, not a statement by the authors.",
          "quote": "for Jetson AGX Orin 64GB Property Mode MAXN 15W 30W* 50W"
        },
        {
          "id": "ev-openpi-libero-inputs",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/src/openpi/policies/libero_policy.py",
          "date": "2026-08-24",
          "verified_on": "2026-10-24",
          "summary": "openpi LIBERO input mapping: one third-person image and one wrist image; the third π slot is zero-padded.",
          "quote": "\"right_wrist_0_rgb\": np.zeros_like(base_image),"
        }
      ]
    },
    {
      "id": "pi05-jetson-agx-orin-64gb-llama-cpp-fp16-2cam-eaglevla",
      "policy": "pi05",
      "compute": "jetson-agx-orin-64gb",
      "checkpoint": "π0.5 LIBERO (openpi pi05_libero) converted to GGUF",
      "date": "2026-10-08",
      "runtime": "llama-cpp",
      "runtime_note": "EagleVLA-Edge: llama.cpp-based VLA engine (GGUF FP16 weights) with CUDA graph reuse, GPU-resident intermediate buffers / KV and unrolled flow-matching denoising.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "224x224",
      "inference_latency_ms": {
        "p50": 374.5
      },
      "inference_hz": 2.67,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin, 'Orin-50W' power mode (AGX Orin 64GB inferred, see ev-orin-board-inferred)",
        "power_mode": "50W (paper label \"Orin-50W\")"
      },
      "note": "Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). Paper's 'control frequency' (6.93 Hz) is 1/reaction time with confidence-based scheduling that skips the VLM; it is not executed actions/s and is not recorded as control_rate_hz. Naive PyTorch π0.5 on the same board per the paper: 1420.8 ms (0.70 Hz) (not filed as a separate run: the baseline precision is not stated).",
      "source_ref": "ev-eaglevla-paper",
      "evidence": [
        {
          "id": "ev-eaglevla-paper",
          "tier": "community-reported",
          "type": "paper",
          "url": "https://arxiv.org/abs/2607.12659v6",
          "date": "2026-10-08",
          "verified_on": "2026-10-24",
          "versions": {
            "arxiv": "2607.12659v6"
          },
          "summary": "EagleVLA (formerly Jetson-PI) v6 Table 4: π0.5 latency on Jetson Orin, LIBERO, with graph reuse + intermediate buffer & unroll; total 374.5 ms.",
          "quote": "+Intermediate Buffer & Unroll 71.9 194.8 107.8 374.5 144.3 6.93"
        },
        {
          "id": "ev-eaglevla-edge-readme",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/PKU-SEC-Lab/EagleVLA-Edge/blob/619e92e878a25df3e9cc0ce87b8f3a61283146e7/README.md",
          "date": "2026-10-23",
          "verified_on": "2026-10-24",
          "summary": "EagleVLA-Edge README reports the same configuration as P50 latency (ms).",
          "quote": "| + Intermediate buffer, GPU KV, and unroll | 71.9 | 194.8 | 107.8 | 374.5 |"
        },
        {
          "id": "ev-orin-board-inferred",
          "tier": "inferred",
          "type": "spec-derivation",
          "url": "https://docs.nvidia.com/jetson/archives/r36.4.3/DeveloperGuide/SD/PlatformPowerAndPerformance/JetsonOrinNanoSeriesJetsonOrinNxSeriesAndJetsonAgxOrinSeries.html",
          "verified_on": "2026-10-24",
          "summary": "INFERRED board identity: the paper names the device only as 'Jetson Orin' and labels this power setting 'Orin-50W' (Table 1). In NVIDIA's Orin power-mode tables only the AGX Orin 64GB has a 50W mode (AGX Orin 32GB: 15/30/40W; Orin NX/Nano: up to 40W/25W). The paper's own hardware table lists both 'Orin 32GB' and 'Orin 64GB', so this mapping is a deduction, not a statement by the authors.",
          "quote": "for Jetson AGX Orin 64GB Property Mode MAXN 15W 30W* 50W"
        },
        {
          "id": "ev-openpi-libero-inputs",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/src/openpi/policies/libero_policy.py",
          "date": "2026-08-24",
          "verified_on": "2026-10-24",
          "summary": "openpi LIBERO input mapping: one third-person image and one wrist image; the third π slot is zero-padded.",
          "quote": "\"right_wrist_0_rgb\": np.zeros_like(base_image),"
        }
      ]
    },
    {
      "id": "pi05-jetson-agx-thor-llama-cpp-fp16-2cam-eaglevla-pr6",
      "policy": "pi05",
      "compute": "jetson-agx-thor",
      "checkpoint": "π0.5 LIBERO (openpi pi05_libero) converted to GGUF",
      "date": "2026-10-23",
      "runtime": "llama-cpp",
      "runtime_note": "EagleVLA-Edge: llama.cpp-based VLA engine (GGUF FP16 weights) with CUDA graph reuse, GPU-resident intermediate buffers / KV and unrolled flow-matching denoising.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "224x224",
      "inference_latency_ms": {
        "p50": 132.8
      },
      "inference_hz": 7.53,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Thor (paper/README: \"Jetson Thor\")"
      },
      "note": "Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). Later configuration from the EagleVLA-Edge README (PR #6); the README notes an accuracy caveat for the two-switch GEMM preset on Thor. Naive PyTorch π0.5 on the same board per the paper: 457.9 ms (2.18 Hz) (not filed as a separate run: the baseline precision is not stated).",
      "source_ref": "ev-eaglevla-edge-readme",
      "evidence": [
        {
          "id": "ev-eaglevla-edge-readme",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/PKU-SEC-Lab/EagleVLA-Edge/blob/619e92e878a25df3e9cc0ce87b8f3a61283146e7/README.md",
          "date": "2026-10-23",
          "verified_on": "2026-10-24",
          "summary": "EagleVLA-Edge README P50 latency table for π0.5 on Jetson Thor: 132.8 ms total.",
          "quote": "| + time cache, and GEMM tuning (PR #6) | **29.0** | **54.9** | **48.9** | **132.8** |"
        },
        {
          "id": "ev-openpi-libero-inputs",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/src/openpi/policies/libero_policy.py",
          "date": "2026-08-24",
          "verified_on": "2026-10-24",
          "summary": "openpi LIBERO input mapping: one third-person image and one wrist image; the third π slot is zero-padded.",
          "quote": "\"right_wrist_0_rgb\": np.zeros_like(base_image),"
        }
      ]
    },
    {
      "id": "pi05-jetson-agx-thor-llama-cpp-fp16-2cam-eaglevla",
      "policy": "pi05",
      "compute": "jetson-agx-thor",
      "checkpoint": "π0.5 LIBERO (openpi pi05_libero) converted to GGUF",
      "date": "2026-10-08",
      "runtime": "llama-cpp",
      "runtime_note": "EagleVLA-Edge: llama.cpp-based VLA engine (GGUF FP16 weights) with CUDA graph reuse, GPU-resident intermediate buffers / KV and unrolled flow-matching denoising.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "224x224",
      "inference_latency_ms": {
        "p50": 202.8
      },
      "inference_hz": 4.93,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Thor (paper/README: \"Jetson Thor\")"
      },
      "note": "Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). Paper's 'control frequency' (10.25 Hz) is 1/reaction time with VLM-skipping scheduling; not executed actions/s. Board inferred to be the 128 GB T5000 class from the paper's 130 W max-power entry for Thor (T4000 is 40–70 W). Naive PyTorch π0.5 on the same board per the paper: 457.9 ms (2.18 Hz) (not filed as a separate run: the baseline precision is not stated).",
      "source_ref": "ev-eaglevla-paper",
      "evidence": [
        {
          "id": "ev-eaglevla-paper",
          "tier": "community-reported",
          "type": "paper",
          "url": "https://arxiv.org/abs/2607.12659v6",
          "date": "2026-10-08",
          "verified_on": "2026-10-24",
          "versions": {
            "arxiv": "2607.12659v6"
          },
          "summary": "EagleVLA (formerly Jetson-PI) v6 Table 4: π0.5 latency on Jetson Thor, LIBERO, with graph reuse + intermediate buffer & unroll; total 202.8 ms.",
          "quote": "+Intermediate Buffer & Unroll 32.0 91.3 79.5 202.8 97.6 10.25"
        },
        {
          "id": "ev-eaglevla-edge-readme",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/PKU-SEC-Lab/EagleVLA-Edge/blob/619e92e878a25df3e9cc0ce87b8f3a61283146e7/README.md",
          "date": "2026-10-23",
          "verified_on": "2026-10-24",
          "summary": "EagleVLA-Edge README reports the same configuration as P50 latency (ms).",
          "quote": "| + Intermediate buffer, GPU KV, and unroll | 32.0 | 91.3 | 79.5 | 202.8 |"
        },
        {
          "id": "ev-openpi-libero-inputs",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/src/openpi/policies/libero_policy.py",
          "date": "2026-08-24",
          "verified_on": "2026-10-24",
          "summary": "openpi LIBERO input mapping: one third-person image and one wrist image; the third π slot is zero-padded.",
          "quote": "\"right_wrist_0_rgb\": np.zeros_like(base_image),"
        }
      ]
    },
    {
      "id": "pi05-jetson-orin-nano-8gb-tensorrt-fp16-2cam-eetmie-libero",
      "policy": "pi05",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "openpi pi05_libero (Physical-Intelligence/openpi)",
      "date": "2026-10-13",
      "runtime": "tensorrt",
      "runtime_note": "Split ONNX export, each graph built into its own TensorRT 10.16.2 engine on the board, engines chained and replayed as one CUDA graph, plus ahead-of-time-compiled Triton kernels (eetmie/jetson-orin-nano-vla). Precision as recorded: FP16 weights and hidden states, FP32 normalization/accumulation/Euler; FP32 accumulation by default.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "640x480",
      "inference_latency_ms": {
        "p50": 438.47,
        "p95": 438.93,
        "mean": 438.45
      },
      "inference_hz": 2.28,
      "action_chunk": 10,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB)",
        "jetpack": "7.2.1",
        "l4t": "R39.2.1",
        "power_mode": "MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz)",
        "tensorrt": "10.16.2.10",
        "os": "Ubuntu 24.04.4 LTS",
        "python": "3.12.3",
        "desktop": "running"
      },
      "note": "Inference cost only; not task success. 300 s run on synthetic 640x480 observations; 10 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 6.39 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing.",
      "source_ref": "ev-eetmie-readme",
      "evidence": [
        {
          "id": "ev-eetmie-readme",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README results table and benchmark conditions (Orin Nano 8 GB Super, TensorRT).",
          "quote": "| [π0.5 LIBERO 3.3B](results/pi05-libero.json) | [openpi](https://github.com/Physical-Intelligence/openpi) `pi05_libero` | 2 | 438.47 ms | 2.28 Hz | 6.39 GB |"
        },
        {
          "id": "ev-eetmie-conditions",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions stated under the table.",
          "quote": "JetPack 7.2.1, MAXN_SUPER with pinned clocks, desktop running (though headless recommended!). 300-second runs, synthetic observations with default denoising steps and views. RAM is whole-board usage during inference."
        },
        {
          "id": "ev-eetmie-results-json",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/results/pi05-libero.json",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Raw run JSON: latency distribution (p50/p95/mean), board state (L4T R39.2.1, MAXN_SUPER, pinned clocks), TensorRT 10.16.2.10, chunk size, steps."
        }
      ]
    },
    {
      "id": "pi05-jetson-orin-nano-8gb-tensorrt-fp16-3cam-eetmie",
      "policy": "pi05",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/pi05_base",
      "date": "2026-10-13",
      "runtime": "tensorrt",
      "runtime_note": "Split ONNX export, each graph built into its own TensorRT 10.16.2 engine on the board, engines chained and replayed as one CUDA graph, plus ahead-of-time-compiled Triton kernels (eetmie/jetson-orin-nano-vla). Precision as recorded: FP16 weights and hidden states, FP32 language residual stream, FP32 normalization/accumulation/Euler; FP32 accumulation by default.",
      "precision": "fp16",
      "cameras": 3,
      "resolution": "640x480",
      "inference_latency_ms": {
        "p50": 643.39,
        "p95": 644.92,
        "mean": 643.31
      },
      "inference_hz": 1.55,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB)",
        "jetpack": "7.2.1",
        "l4t": "R39.2.1",
        "power_mode": "MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz)",
        "tensorrt": "10.16.2.10",
        "os": "Ubuntu 24.04.4 LTS",
        "python": "3.12.3",
        "desktop": "running"
      },
      "note": "Inference cost only; not task success. 300 s run on synthetic 640x480 observations; 10 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 6.18 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing.",
      "source_ref": "ev-eetmie-readme",
      "evidence": [
        {
          "id": "ev-eetmie-readme",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README results table and benchmark conditions (Orin Nano 8 GB Super, TensorRT).",
          "quote": "| [π0.5 base 3.3B](results/pi05-base.json) | [`lerobot/pi05_base`](https://huggingface.co/lerobot/pi05_base) | 3 | 643.39 ms | 1.55 Hz | 6.18 GB |"
        },
        {
          "id": "ev-eetmie-conditions",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions stated under the table.",
          "quote": "JetPack 7.2.1, MAXN_SUPER with pinned clocks, desktop running (though headless recommended!). 300-second runs, synthetic observations with default denoising steps and views. RAM is whole-board usage during inference."
        },
        {
          "id": "ev-eetmie-results-json",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/results/pi05-base.json",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Raw run JSON: latency distribution (p50/p95/mean), board state (L4T R39.2.1, MAXN_SUPER, pinned clocks), TensorRT 10.16.2.10, chunk size, steps."
        }
      ]
    },
    {
      "id": "pi05-rtx-4090-llama-cpp-fp16-2cam-eaglevla-pr6",
      "policy": "pi05",
      "compute": "rtx-4090",
      "checkpoint": "π0.5 LIBERO (openpi pi05_libero) converted to GGUF",
      "date": "2026-10-23",
      "runtime": "llama-cpp",
      "runtime_note": "EagleVLA-Edge: llama.cpp-based VLA engine (GGUF FP16 weights) with CUDA graph reuse, GPU-resident intermediate buffers / KV and unrolled flow-matching denoising.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "224x224",
      "inference_latency_ms": {
        "p50": 51.7
      },
      "inference_hz": 19.34,
      "outcome": "runs",
      "platform": {
        "board": "NVIDIA RTX 4090",
        "ggml_cuda_f16_f32_out": "0 (README preset for RTX 4090)"
      },
      "note": "Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). From the EagleVLA-Edge README RTX 4090 table (P50), PR #6 configuration. The README's naive π0.5 baseline on this GPU: 162.9 ms (not filed as a separate run: the baseline precision is not stated).",
      "source_ref": "ev-eaglevla-edge-readme",
      "evidence": [
        {
          "id": "ev-eaglevla-edge-readme",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/PKU-SEC-Lab/EagleVLA-Edge/blob/619e92e878a25df3e9cc0ce87b8f3a61283146e7/README.md",
          "date": "2026-10-23",
          "verified_on": "2026-10-24",
          "summary": "EagleVLA-Edge README P50 latency table for π0.5 on RTX 4090: 51.7 ms total.",
          "quote": "| + time cache, and GEMM tuning (PR #6) | **10.3** | **21.2** | **20.1** | **51.7** |"
        },
        {
          "id": "ev-openpi-libero-inputs",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/src/openpi/policies/libero_policy.py",
          "date": "2026-08-24",
          "verified_on": "2026-10-24",
          "summary": "openpi LIBERO input mapping: one third-person image and one wrist image; the third π slot is zero-padded.",
          "quote": "\"right_wrist_0_rgb\": np.zeros_like(base_image),"
        }
      ]
    },
    {
      "id": "pi05-rtx-4090-llama-cpp-fp16-2cam-eaglevla",
      "policy": "pi05",
      "compute": "rtx-4090",
      "checkpoint": "π0.5 LIBERO (openpi pi05_libero) converted to GGUF",
      "date": "2026-10-23",
      "runtime": "llama-cpp",
      "runtime_note": "EagleVLA-Edge: llama.cpp-based VLA engine (GGUF FP16 weights) with CUDA graph reuse, GPU-resident intermediate buffers / KV and unrolled flow-matching denoising.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "224x224",
      "inference_latency_ms": {
        "p50": 58.5
      },
      "inference_hz": 17.09,
      "outcome": "runs",
      "platform": {
        "board": "NVIDIA RTX 4090",
        "ggml_cuda_f16_f32_out": "0 (README preset for RTX 4090)"
      },
      "note": "Inference cost only; not task success. Full VLM + action-expert pass (ViT + LLM + 10-step action expert), LIBERO inputs (agent-view + wrist at 224x224). From the EagleVLA-Edge README RTX 4090 table (P50). The README's naive π0.5 baseline on this GPU: 162.9 ms (not filed as a separate run: the baseline precision is not stated).",
      "source_ref": "ev-eaglevla-edge-readme",
      "evidence": [
        {
          "id": "ev-eaglevla-edge-readme",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/PKU-SEC-Lab/EagleVLA-Edge/blob/619e92e878a25df3e9cc0ce87b8f3a61283146e7/README.md",
          "date": "2026-10-23",
          "verified_on": "2026-10-24",
          "summary": "EagleVLA-Edge README P50 latency table for π0.5 on RTX 4090: 58.5 ms total.",
          "quote": "| + Intermediate buffer, GPU KV, and unroll | 10.5 | 22.5 | 25.7 | 58.5 |"
        },
        {
          "id": "ev-openpi-libero-inputs",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/Physical-Intelligence/openpi/blob/215abfb217dbac7d5f1273282331b9b1866c0479/src/openpi/policies/libero_policy.py",
          "date": "2026-08-24",
          "verified_on": "2026-10-24",
          "summary": "openpi LIBERO input mapping: one third-person image and one wrist image; the third π slot is zero-padded.",
          "quote": "\"right_wrist_0_rgb\": np.zeros_like(base_image),"
        }
      ]
    },
    {
      "id": "smolvla-jetson-orin-nano-8gb-onnx-split-fp16-2cam-eetmie-2026-09",
      "policy": "smolvla",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/smolvla_base",
      "date": "2026-09-05",
      "runtime": "onnx-split",
      "runtime_note": "Split ONNX graphs on ONNX Runtime (onnxruntime-gpu 1.24) with the TensorRT execution provider (FP16); superseded in the same repo by the pure-TensorRT runs.",
      "precision": "fp16",
      "cameras": 2,
      "inference_latency_ms": {
        "p50": 189.89,
        "p95": 190.93
      },
      "inference_hz": 5.25,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Super 8 GB developer kit",
        "jetpack": "7.2 (L4T R39.2.x, per the repo host-setup doc at this commit)",
        "power_mode": "MAXN_SUPER, pinned clocks",
        "cuda": "13.2",
        "os": "Ubuntu 24.04",
        "onnxruntime-gpu": "1.24",
        "tensorrt": "10.16.2 (system, via ONNX Runtime TensorRT EP)"
      },
      "note": "Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Configured EP priority 7 graphs TensorRT / 2 CPU. System RAM 2385 MB in use.",
      "source_ref": "ev-eetmie-readme-2026-09",
      "evidence": [
        {
          "id": "ev-eetmie-readme-2026-09",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/README.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README \"Measured fit\" table at commit b6185e1 (PyTorch FP32 vs split ONNX FP16 on Orin Nano 8 GB Super).",
          "quote": "| SmolVLA split ONNX FP16 | 2 | 189.89 ms | 190.93 ms | 5.25 Hz |"
        },
        {
          "id": "ev-eetmie-results-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/RESULTS.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Generated results at the same commit: speed, footprint and control-rate fit per run."
        },
        {
          "id": "ev-eetmie-host-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/01-host-setup.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Board state at that commit: JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.",
          "quote": "JetPack 7.2 is L4T R39.2.0: Ubuntu 24.04, kernel 6.8, Python 3.12, CUDA 13.2, cuDNN 9.20,"
        }
      ]
    },
    {
      "id": "smolvla-jetson-orin-nano-8gb-pytorch-eager-fp32-2cam-eetmie-2026-09",
      "policy": "smolvla",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/smolvla_base",
      "date": "2026-09-05",
      "runtime": "pytorch-eager",
      "precision": "fp32",
      "cameras": 2,
      "inference_latency_ms": {
        "p50": 1167.93,
        "p95": 1176.65
      },
      "inference_hz": 0.86,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Super 8 GB developer kit",
        "jetpack": "7.2 (L4T R39.2.x, per the repo host-setup doc at this commit)",
        "power_mode": "MAXN_SUPER, pinned clocks",
        "cuda": "13.2",
        "os": "Ubuntu 24.04",
        "lerobot": "0.5.1"
      },
      "note": "Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Stock LeRobot SmolVLAPolicy in PyTorch FP32 (LeRobot 0.5.1 env, JetPack-matched torch). RSS peak 3826 MB, system RAM 4126 MB in use.",
      "source_ref": "ev-eetmie-readme-2026-09",
      "evidence": [
        {
          "id": "ev-eetmie-readme-2026-09",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/README.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README \"Measured fit\" table at commit b6185e1 (PyTorch FP32 vs split ONNX FP16 on Orin Nano 8 GB Super).",
          "quote": "| SmolVLA PyTorch FP32 | 2 | 1167.93 ms | 1176.65 ms | 0.86 Hz |"
        },
        {
          "id": "ev-eetmie-results-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/RESULTS.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Generated results at the same commit: speed, footprint and control-rate fit per run."
        },
        {
          "id": "ev-eetmie-host-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/01-host-setup.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Board state at that commit: JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.",
          "quote": "JetPack 7.2 is L4T R39.2.0: Ubuntu 24.04, kernel 6.8, Python 3.12, CUDA 13.2, cuDNN 9.20,"
        }
      ]
    },
    {
      "id": "smolvla-jetson-orin-nano-8gb-tensorrt-fp16-2cam-eetmie",
      "policy": "smolvla",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/smolvla_base",
      "date": "2026-10-13",
      "runtime": "tensorrt",
      "runtime_note": "Split ONNX export, each graph built into its own TensorRT 10.16.2 engine on the board, engines chained and replayed as one CUDA graph, plus ahead-of-time-compiled Triton kernels (eetmie/jetson-orin-nano-vla). Precision as recorded: mixed fp16 (strongly typed); FP32 accumulation by default.",
      "precision": "fp16",
      "cameras": 2,
      "resolution": "640x480",
      "inference_latency_ms": {
        "p50": 107.55,
        "p95": 108.48,
        "mean": 107.53
      },
      "inference_hz": 9.3,
      "action_chunk": 50,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB)",
        "jetpack": "7.2.1",
        "l4t": "R39.2.1",
        "power_mode": "MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz)",
        "tensorrt": "10.16.2.10",
        "os": "Ubuntu 24.04.4 LTS",
        "python": "3.12.3",
        "desktop": "running"
      },
      "note": "Inference cost only; not task success. 300 s run on synthetic 640x480 observations resized by the exported policy to 512x512; 10 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 1.78 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing.",
      "source_ref": "ev-eetmie-readme",
      "evidence": [
        {
          "id": "ev-eetmie-readme",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README results table and benchmark conditions (Orin Nano 8 GB Super, TensorRT).",
          "quote": "| [SmolVLA 450M](results/smolvla-base.json) | [`lerobot/smolvla_base`](https://huggingface.co/lerobot/smolvla_base) | 2 | 107.55 ms | 9.30 Hz | 1.78 GB |"
        },
        {
          "id": "ev-eetmie-conditions",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions stated under the table.",
          "quote": "JetPack 7.2.1, MAXN_SUPER with pinned clocks, desktop running (though headless recommended!). 300-second runs, synthetic observations with default denoising steps and views. RAM is whole-board usage during inference."
        },
        {
          "id": "ev-eetmie-results-json",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/results/smolvla-base.json",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Raw run JSON: latency distribution (p50/p95/mean), board state (L4T R39.2.1, MAXN_SUPER, pinned clocks), TensorRT 10.16.2.10, chunk size, steps."
        }
      ]
    },
    {
      "id": "xvla-jetson-orin-nano-8gb-onnx-split-fp16-3cam-eetmie-2026-09",
      "policy": "xvla",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/xvla-base",
      "date": "2026-09-05",
      "runtime": "onnx-split",
      "runtime_note": "Split ONNX graphs on ONNX Runtime (onnxruntime-gpu 1.24) with the TensorRT execution provider (FP16); superseded in the same repo by the pure-TensorRT runs.",
      "precision": "fp16",
      "cameras": 3,
      "inference_latency_ms": {
        "p50": 391.55,
        "p95": 407.33
      },
      "inference_hz": 2.55,
      "action_chunk": 30,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Super 8 GB developer kit",
        "jetpack": "7.2 (L4T R39.2.x, per the repo host-setup doc at this commit)",
        "power_mode": "MAXN_SUPER, pinned clocks",
        "cuda": "13.2",
        "os": "Ubuntu 24.04",
        "onnxruntime-gpu": "1.24",
        "tensorrt": "10.16.2 (system, via ONNX Runtime TensorRT EP)"
      },
      "note": "Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Configured EP priority 12 graphs TensorRT. System RAM 5395 MB in use.",
      "source_ref": "ev-eetmie-readme-2026-09",
      "evidence": [
        {
          "id": "ev-eetmie-readme-2026-09",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/README.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README \"Measured fit\" table at commit b6185e1 (PyTorch FP32 vs split ONNX FP16 on Orin Nano 8 GB Super).",
          "quote": "| X-VLA split ONNX FP16 | 3 | 391.55 ms | 407.33 ms | 2.55 Hz |"
        },
        {
          "id": "ev-eetmie-results-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/RESULTS.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Generated results at the same commit: speed, footprint and control-rate fit per run."
        },
        {
          "id": "ev-eetmie-host-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/01-host-setup.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Board state at that commit: JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.",
          "quote": "JetPack 7.2 is L4T R39.2.0: Ubuntu 24.04, kernel 6.8, Python 3.12, CUDA 13.2, cuDNN 9.20,"
        }
      ]
    },
    {
      "id": "xvla-jetson-orin-nano-8gb-pytorch-eager-fp32-3cam-eetmie-2026-09",
      "policy": "xvla",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/xvla-base",
      "date": "2026-09-05",
      "runtime": "pytorch-eager",
      "precision": "fp32",
      "cameras": 3,
      "inference_latency_ms": {
        "p50": 2313.5,
        "p95": 2320.89
      },
      "inference_hz": 0.43,
      "action_chunk": 30,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Super 8 GB developer kit",
        "jetpack": "7.2 (L4T R39.2.x, per the repo host-setup doc at this commit)",
        "power_mode": "MAXN_SUPER, pinned clocks",
        "cuda": "13.2",
        "os": "Ubuntu 24.04",
        "lerobot": "0.6.1"
      },
      "note": "Inference cost only; not task success. Deterministic synthetic observations, pinned MAXN_SUPER clocks. Stock LeRobot X-VLA policy in PyTorch FP32 (LeRobot 0.6.1 env). At 30 Hz control the 30-step chunk is overrun (author's control-rate fit). System RAM 5450 MB in use.",
      "source_ref": "ev-eetmie-readme-2026-09",
      "evidence": [
        {
          "id": "ev-eetmie-readme-2026-09",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/README.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README \"Measured fit\" table at commit b6185e1 (PyTorch FP32 vs split ONNX FP16 on Orin Nano 8 GB Super).",
          "quote": "| X-VLA PyTorch FP32 | 3 | 2313.50 ms | 2320.89 ms | 0.43 Hz |"
        },
        {
          "id": "ev-eetmie-results-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/RESULTS.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Generated results at the same commit: speed, footprint and control-rate fit per run."
        },
        {
          "id": "ev-eetmie-host-2026-09",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/b6185e17c5006e449a5d2b1c1a31ae7f26ce9f9f/docs/01-host-setup.md",
          "date": "2026-09-05",
          "verified_on": "2026-10-24",
          "summary": "Board state at that commit: JetPack 7.2 / L4T R39.2, CUDA 13.2, TensorRT 10.16.2.",
          "quote": "JetPack 7.2 is L4T R39.2.0: Ubuntu 24.04, kernel 6.8, Python 3.12, CUDA 13.2, cuDNN 9.20,"
        }
      ]
    },
    {
      "id": "xvla-jetson-orin-nano-8gb-tensorrt-fp16-3cam-eetmie",
      "policy": "xvla",
      "compute": "jetson-orin-nano-8gb",
      "checkpoint": "lerobot/xvla-base",
      "date": "2026-10-13",
      "runtime": "tensorrt",
      "runtime_note": "Split ONNX export, each graph built into its own TensorRT 10.16.2 engine on the board, engines chained and replayed as one CUDA graph, plus ahead-of-time-compiled Triton kernels (eetmie/jetson-orin-nano-vla). Precision as recorded: mixed fp16 (strongly typed); FP32 accumulation by default.",
      "precision": "fp16",
      "cameras": 3,
      "resolution": "640x480",
      "inference_latency_ms": {
        "p50": 342.59,
        "p95": 343.47,
        "mean": 342.53
      },
      "inference_hz": 2.92,
      "action_chunk": 30,
      "outcome": "runs",
      "platform": {
        "board": "Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB)",
        "jetpack": "7.2.1",
        "l4t": "R39.2.1",
        "power_mode": "MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz)",
        "tensorrt": "10.16.2.10",
        "os": "Ubuntu 24.04.4 LTS",
        "python": "3.12.3",
        "desktop": "running"
      },
      "note": "Inference cost only; not task success. 300 s run on synthetic 640x480 observations resized by the exported policy to 224x224; 10 denoising steps; latency is runtime-only over a pre-materialized observation ring. Whole-board RAM in use 3.04 GB (unified memory, desktop running). Actions were gated against the stock policy (cosine >= 0.99999) before timing.",
      "source_ref": "ev-eetmie-readme",
      "evidence": [
        {
          "id": "ev-eetmie-readme",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "eetmie/jetson-orin-nano-vla README results table and benchmark conditions (Orin Nano 8 GB Super, TensorRT).",
          "quote": "| [X-VLA 0.9B](results/xvla-base.json) | [`lerobot/xvla-base`](https://huggingface.co/lerobot/xvla-base) | 3 | 342.59 ms | 2.92 Hz | 3.04 GB |"
        },
        {
          "id": "ev-eetmie-conditions",
          "tier": "community-reported",
          "type": "community-post",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/README.md",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Benchmark conditions stated under the table.",
          "quote": "JetPack 7.2.1, MAXN_SUPER with pinned clocks, desktop running (though headless recommended!). 300-second runs, synthetic observations with default denoising steps and views. RAM is whole-board usage during inference."
        },
        {
          "id": "ev-eetmie-results-json",
          "tier": "community-reported",
          "type": "source-code",
          "url": "https://github.com/eetmie/jetson-orin-nano-vla/blob/6e2d857ca5e3342fb2e29afd0cb60e295c10cdf5/results/xvla-base.json",
          "date": "2026-10-13",
          "verified_on": "2026-10-24",
          "summary": "Raw run JSON: latency distribution (p50/p95/mean), board state (L4T R39.2.1, MAXN_SUPER, pinned clocks), TensorRT 10.16.2.10, chunk size, steps."
        }
      ]
    }
  ]
}