GR00T N1.7
NVIDIA's open cross-embodiment foundation model for humanoid and manipulation skills: a Cosmos-Reason2-2B (Qwen3-VL architecture) vision-language backbone feeding a flow-matching diffusion-transformer (DiT) action head that denoises continuous action chunks conditioned on vision, language and proprioception.
3B paramsopen weightsRTCverified 2026-10-24paper →
Architecture
- Family
- vla
- Parameters
- 3,000,000,000 (Model card states 3,000,000,000; the Isaac-GR00T README calls it the 'Base model (3B params)'. Both are rounded figures.)src
- Backbone
- Cosmos-Reason2-2B (Qwen3-VL architecture) VLM with SigLip2 vision transformersrc
- Action head
- flow-matchingsrc
- Action chunking
- yes
- Real-Time Chunking
- supportedsrc
- Code license
- Apache-2.0
- Weights license
- NVIDIA Open Model License
Compute requirements
| Memory | Precision | Conditions | |
|---|---|---|---|
| train | ≥24 GB · ≤40 GBsrc | — | Peak training VRAM at batch size 8 with AdamW (LeRobot default), "Multimodal" group (`groot`, `eo1`) in the LeRobot compute hardware guide; suggested starter GPU "A100 40 GB+". The guide calls its figures indicative, order-of-magnitude numbers for picking hardware, not exact predictions. Precision not stated. · batch 8 |
| inference | ≥16 GBsrc | — | Upstream Isaac-GR00T N1.7 README hardware requirement for inference with NVIDIA's own gr00t package: 1 GPU with 16 GB+ VRAM (e.g. RTX 4090, L40, H100, Jetson AGX Thor/Orin, DGX Spark). Not a LeRobot measurement; precision and batch size not stated. The same README recommends 40 GB+ GPUs for fine-tuning. |
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 14 runs · inference cost only, not task success
| Compute | Runtime | Precision | Cams | p50 / p95 ms | Inference | Platform | Source |
|---|---|---|---|---|---|---|---|
| H100 80GB | TensorRT | bf16 | 2 @256x256 | 22.9 / — | 43.67 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| RTX PRO 6000 Blackwell Server | TensorRT | bf16 | 2 @256x256 | 24.7 / — | 40.49 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| L40 | TensorRT | bf16 | 2 @256x256 | 35 / — | 28.57 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| Jetson AGX Thor 128GB | TensorRT | mixed | 2 @256x256 | 35.6 / — | 28.09 Hz | jetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| RTX PRO 6000 Blackwell Server | torch.compile | bf16 | 2 @256x256 | 52.1 / — | 19.19 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| H100 80GB | torch.compile | bf16 | 2 @256x256 | 53.4 / — | 18.73 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| L40 | torch.compile | bf16 | 2 @256x256 | 66.8 / — | 14.97 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| RTX PRO 6000 Blackwell Server | PyTorch eager | bf16 | 2 @256x256 | 75.7 / — | 13.21 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| Jetson AGX Thor 128GB | TensorRT | bf16 | 2 @256x256 | 76.8 / — | 13.02 Hz | jetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| Jetson AGX Thor 128GB | torch.compile | bf16 | 2 @256x256 | 100.5 / — | 9.95 Hz | jetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| H100 80GB | PyTorch eager | bf16 | 2 @256x256 | 101.5 / — | 9.85 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| Jetson AGX Thor 128GB | PyTorch eager | bf16 | 2 @256x256 | 106.8 / — | 9.36 Hz | jetpack 7.2 · power_mode MAXN with jetson_clocks · cuda 13.2 · torch 2.13.0conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| L40 | PyTorch eager | bf16 | 2 @256x256 | 123.7 / — | 8.08 Hz | —conditionstorch_threads: 8 intra-op / 1 inter-op 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). | T3 · community reported |
| Jetson Orin Nano 8GB (Super) | TensorRT | fp16 | 3 @640x480 | 273.7 / 275.4 | 3.66 Hz | jetpack 7.2.1 · power_mode MAXN_SUPER, clocks pinned with jetson_clocks (GPU 1020 MHz, CPU 1728 MHz) · tensorrt 10.16.2.10conditionsboard: Jetson Orin Nano Engineering Reference Developer Kit Super (8 GB) · l4t: R39.2.1 · os: Ubuntu 24.04.4 LTS · python: 3.12.3 · desktop: running 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). | T3 · community reported |
Robots running GR00T N1.7 11 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| Bimanual SO-100/SO-101 | partial | T3 · community reported | 2026-10-25 today | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| Franka Emika Panda | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | NVIDIA documents the DROID deployment path (inference server + DROID robot control script) but publishes no on-robot DROID success rates for N1.7. The README reports only offline MSE on demo data. |
| I2RT YAM | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| ROBOTIS OMX | works with caveats | T3 · community reported | 2026-10-24 1d ago | Community-reported only; not documented upstream or maintainer-tested. |
| OpenArm | works with caveats | T1 · upstream documented | 2026-10-25 today | Documented only in Enactic's fork of Isaac-GR00T (branch n1.7-release-ga-openarm, a single commit ahead of NVIDIA main as of 2026-10-23), not in NVIDIA's Isaac-GR00T or in LeRobot. |
| reBot Arm B601-DM | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Documented in NVIDIA's Isaac-GR00T repository, not through LeRobot's groot policy: the real-robot client uses Seeed's LeRobot fork plus the lerobot-robot-seeed-b601 plugin (--robot.type=seeed_b601_dm_follower), not LeRobot's native rebot_b601_follower. |
| SO-100 | works | T1 · upstream documented | 2026-10-24 1d ago | |
| SO-101 | works | T1 · upstream documented | 2026-10-24 1d ago | |
| Trossen WidowX AI | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| Unitree G1 | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Documented in NVIDIA's Isaac-GR00T repository (embodiment tag UNITREE_G1_SONIC), not through LeRobot: whole-body motion is executed by the GR00T Whole-Body Control / GEAR-SONIC robot-side controller, with GR00T N1.7 served from the GR00T policy server. |
| WidowX 250 S | partial | T1 · upstream documented | 2026-10-25 today | Simulation only: SimplerEnv WidowX + Bridge (N1.7 post-trained on BridgeData V2); no real-robot training or inference documented. |
No verified evidence, but both are available in LeRobot (ecosystem context): AgileX PiPER, ARX5, EarthRover Mini Plus, HopeJR, Koch v1.1, LeKiwi, Reachy 2, Universal Robots UR5e, UFACTORY xArm, XLeRobot.
Framework support
| Framework | Support | Since | Type string | Docs |
|---|---|---|---|---|
| LeRobot | nativesrc | 0.6.0 | groot | docs |
| Isaac GR00T | upstream-reposrc | n1.7-release | — | repo |
| LeRobot: `groot` is GR00T N1.7 from v0.6.0. LeRobot v0.4.0–v0.5.1 shipped GR00T N1.5 under the same `groot` type (default base model nvidia/GR00T-N1.5-3B); N1.5 was removed in v0.6.0 and N1.5 checkpoints are rejected. To keep N1.5, pin lerobot==0.5.1. groot.mdx shows rollout with --inference.type=rtc. | ||||
| Isaac GR00T: N1.7 is an Early Access release (tag n1.7-release, 2026-04-18); older generations live on the n1.6-release / n1.5-release tags. | ||||
Checkpoints
- nvidia/GR00T-N1.7-3B · Base model; NVIDIA Open Model License.
- nvidia/gr00t17-lerobot-libero_spatial-640 · LeRobot-trained LIBERO Spatial checkpoint (95% in LeRobot docs).
- nvidia/gr00t17-lerobot-libero_object-640
- nvidia/gr00t17-lerobot-libero_goal-640
- nvidia/gr00t17-lerobot-libero_10-640
- nvidia/GR00T-N1.7-LIBERO · Upstream (Isaac-GR00T format) LIBERO checkpoints.
Related terms
Action chunking · Action horizon · Embodiment gap · Flow matching · Proprioception · Vision-language-action model
Sources
- T1 · upstream documentedmodel-cardverified 2026-10-24 (1d ago)
N1.7 model card gives the parameter count and the flow-matching DiT action head over SigLip2 vision features.
“**Number of Model Parameters:** 3,000,000,000”
https://huggingface.co/nvidia/GR00T-N1.7-3B - T1 · upstream documentedmodel-cardverified 2026-10-24 (1d ago)
Model card describes the flow matching action transformer that models a chunk of actions.
“uses a flow matching action transformer to model a chunk of actions conditioned on vision, language and proprioception.”
https://huggingface.co/nvidia/GR00T-N1.7-3B - T1 · upstream documentedmodel-cardverified 2026-10-24 (1d ago)
Model card describes proprioception encoding and padding.
“To handle variable-dimension proprioception, inputs are padded to a configurable max length before feeding into the MLP.”
https://huggingface.co/nvidia/GR00T-N1.7-3B - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot integrates GR00T N1.7 only; N1.5 was removed and needs lerobot==0.5.1.
“To keep using an N1.5 checkpoint, pin the last release that supports it: `pip install 'lerobot==0.5.1'`.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/groot.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot GR00T docs deploy with RTC inference via lerobot-rollout.
“--inference.rtc.enabled=True \ # set to False if it causes inference instability”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/groot.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `groot`.
“@PreTrainedConfig.register_subclass("groot")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/groot/configuration_groot.py - T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
v0.6.0 replaced GR00T N1.5 with N1.7 (PR
“GR00T N1.5 is replaced by N1.7 (pin `lerobot==0.5.1` if you need N1.5).”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0 - T1 · upstream documentedrelease-notessource 2025-10-23verified 2026-10-24 (1d ago)lerobot 0.4.0
The `groot` policy (then GR00T N1.5) first shipped in v0.4.0.
“feat(policies): add Nvidia Gr00t N1.5 model by @imstevenpmwork”
https://github.com/huggingface/lerobot/releases/tag/v0.4.0 - T1 · upstream documentedofficial-docssource 2026-04-18verified 2026-10-24 (1d ago)
Isaac-GR00T n1.7-release README describes N1.7 as an Early Access release with a Cosmos-Reason2-2B / Qwen3-VL backbone.
“We just released GR00T N1.7 Early Access, the latest version of GR00T N1 with a new VLM backbone (Cosmos-Reason2-2B / Qwen3-VL) and improved performance.”
https://github.com/NVIDIA/Isaac-GR00T/blob/n1.7-release/README.md - T1 · upstream documentedofficial-docssource 2026-04-18verified 2026-10-24 (1d ago)
Upstream hardware requirements for N1.7 inference (16 GB+) and fine-tuning (40 GB+ recommended).
“**Inference:** 1 GPU with 16 GB+ VRAM (e.g., RTX 4090, L40, H100, Jetson AGX Thor/Orin, DGX Spark).”
https://github.com/NVIDIA/Isaac-GR00T/blob/n1.7-release/README.md - T1 · upstream documentedofficial-docssource 2026-04-18verified 2026-10-24 (1d ago)
Code is Apache 2.0; model weights are under the NVIDIA Open Model License.
“- **Model weights:** [NVIDIA Open Model License](https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-open-model-license/)”
https://github.com/NVIDIA/Isaac-GR00T/blob/n1.7-release/README.md - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Compute hardware guide: Multimodal group (`groot`, `eo1`) peak VRAM ~24–40GB at BS 8 with AdamW; figures are indicative.
“| Multimodal | `groot`, `eo1` | ~24–40GB | A100 40 GB+ |”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/hardware_guide.mdx