Evo-1
Lightweight vision-language-action model from MINT-SJTU: an InternVL3-1B backbone (14 language layers kept) with a continuous flow-matching action head predicting 50-step action chunks from one to three camera views, state and a language task. Trained in two stages (action head only, then VLM fine-tuning).
770M paramsopen weightsRTCverified 2026-10-24paper →
Architecture
- Family
- vla
- Parameters
- 770,000,000 (Authors state 0.77B parameters (rounded).)src
- Backbone
- InternVL3-1B (OpenGVLab/InternVL3-1B-hf), 14 language layerssrc
- Action head
- flow-matchingsrc
- Action chunking
- yes
- Real-Time Chunking
- supportedsrc
- Code license
- MIT (upstream MINT-SJTU/Evo-1); LeRobot integration Apache-2.0
- Cameras
- min 1 · max 3src
- Language input
- required
Compute requirements
No sourced requirements yet.
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 3 runs · inference cost only, not task success
| Compute | Runtime | Precision | Cams | p50 / p95 ms | Inference | Platform | Source |
|---|---|---|---|---|---|---|---|
| Jetson Orin Nano 8GB (Super) | ONNX (split) | fp16 | 1 | 289.18 / 292.12 | 3.45 Hz | 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 · tensorrt 10.16.2 (system, via ONNX Runtime TensorRT EP)conditionsboard: Jetson Orin Nano Super 8 GB developer kit · os: Ubuntu 24.04 · onnxruntime-gpu: 1.24 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". | T3 · community reported |
| Jetson Orin Nano 8GB (Super) | TensorRT | fp16 | 2 @640x480 | 332.77 / 333.63 | 3 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 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. | T3 · community reported |
| Jetson Orin Nano 8GB (Super) | ONNX (split) | fp16 | 2 | 414.67 / 424.72 | 2.41 Hz | 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 · tensorrt 10.16.2 (system, via ONNX Runtime TensorRT EP)conditionsboard: Jetson Orin Nano Super 8 GB developer kit · os: Ubuntu 24.04 · onnxruntime-gpu: 1.24 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. | T3 · community reported |
Robots running Evo-1 6 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| ALOHA (stationary) | partial | T1 · upstream documented | 2026-10-24 1d ago | The upstream Evo-1 repo ships a community-contributed real-ALOHA inference client (ROS1 Noetic, Interbotix); no ALOHA training recipe, checkpoint or on-robot results are documented. |
| Franka Emika Panda | partial | T1 · upstream documented | 2026-10-25 today | Simulation only: LIBERO (simulated Franka Panda); no real-robot training or inference documented. |
| ROBOTIS OMX | works with caveats | T1 · upstream documented | 2026-10-25 today | Physical-run evidence is post-0.6.1, unreleased LeRobot main (remote-inference docs, PR #4937); no released LeRobot version documents it. |
| SO-100 | works | T1 · upstream documented | 2026-10-24 1d ago | |
| SO-101 | works | T1 · upstream documented | 2026-10-24 1d ago | |
| UFACTORY xArm | works with caveats | T1 · upstream documented | 2026-10-24 1d ago | Evidence covers the xArm6 only, driven by Evo-1's own websocket client script (official xArmAPI), not by the LeRobot xArm plugin; no xArm checkpoint is published. |
No verified evidence, but both are available in LeRobot (ecosystem context): AgileX PiPER, ARX5, Bimanual SO-100/SO-101, EarthRover Mini Plus, HopeJR, I2RT YAM, Koch v1.1, LeKiwi, OpenArm, Reachy 2, reBot Arm B601-DM, Trossen WidowX AI, Unitree G1, Universal Robots UR5e, XLeRobot.
Framework support
Checkpoints
- zuoxingdong/evo1_libero · LeRobot-format Stage-2 LIBERO checkpoint referenced by the LeRobot docs (96.65% LIBERO average at step 70k, single run).
- MINT-SJTU/Evo1_LIBERO · Upstream (non-LeRobot-format) LIBERO weights from the Evo-1 repo.
- MINT-SJTU/Evo1_MetaWorld · Upstream (non-LeRobot-format) Meta-World weights from the Evo-1 repo.
Related terms
Sources
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot EVO1 docs describe an InternVL3 backbone with a continuous flow-matching action head.
“EVO1 is a Vision-Language-Action policy for robot control built around an InternVL3 backbone and a continuous flow-matching action head.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/evo1.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
EVO1 expects one to `policy.max_views` (default 3) cameras, observation.state, action and a language task.
“One to `policy.max_views` visual observations, for example `observation.images.image`”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/evo1.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
EVO1 rollout example enables RTC inference as an option.
“--inference.type=rtc \ # optional”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/evo1.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `evo1`; default max_views 3, max_state_dim 24, VLM OpenGVLab/InternVL3-1B-hf.
“@PreTrainedConfig.register_subclass("evo1")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/evo1/configuration_evo1.py - T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
EVO1 was added in LeRobot v0.6.0.
“* feat(policies): add EVO1 policy by @imstevenpmwork in https://github.com/huggingface/lerobot/pull/3908”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0 - T1 · upstream documentedpapersource 2025-11-06verified 2026-10-24 (1d ago)
Evo-1 paper states 0.77B parameters.
“We propose Evo-1, a lightweight VLA architecture with only 0.77B parameters”
https://arxiv.org/abs/2511.04555v1 - T3 · community reportedsource-codesource 2026-08-18verified 2026-10-24 (1d ago)
GitHub API reports the upstream MINT-SJTU/Evo-1 repo license as MIT; README links the MINT-SJTU/Evo1_LIBERO and Evo1_MetaWorld weights.
https://github.com/MINT-SJTU/Evo-1/tree/5fd14b015013c4fd0aacf5f8f48f868ca9b870a2 - T3 · community reportedhub-checkpointverified 2026-10-24 (1d ago)
Hub API: zuoxingdong/evo1_libero (license tag 'other'), MINT-SJTU/Evo1_LIBERO and MINT-SJTU/Evo1_MetaWorld (no license tag) exist.
https://huggingface.co/zuoxingdong/evo1_libero