EO-1
Unified embodied foundation model from the EO-Robotics project: a Qwen2.5-VL-3B backbone trained on interleaved vision-text-action data. LeRobot integrates the robot-control path: a continuous flow-matching action head that predicts action chunks (default 8 steps) from images, EO-1 state tokens and a language task.
3B paramsopen weightsverified 2026-10-24paper →
Architecture
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 | ≥6.5 GBsrc | — | Upstream EO-1 repository README: loading the pretrained model with transformers + lerobot and running inference "requires ~6.5GB GPU memory". Refers to the upstream EO-1 code path, not measured on the LeRobot v0.6.1 port; precision and batch size not stated. |
Training requirements: see also the LeRobot compute hardware guide, the authoritative source for LeRobot training envelopes.
Measured inference 0 runs · inference cost only, not task success
No measured runs yet. Submit one with its runtime, precision and camera count.
Robots running EO-1 5 of 24 with evidence
| Robot | State | Tier | Verified | Detail |
|---|---|---|---|---|
| Franka Emika Panda | works | T1 · upstream documented | 2026-10-24 1d ago | |
| ROBOTIS OMX | partial | T3 · community reported | 2026-10-25 today | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| OpenArm | partial | T3 · community reported | 2026-10-24 1d ago | Training evidenced by public checkpoint(s); on-robot inference not documented by a primary source. |
| SO-100 | works | T1 · upstream documented | 2026-10-24 1d ago | |
| WidowX 250 S | works | T1 · upstream documented | 2026-10-24 1d ago |
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, Reachy 2, reBot Arm B601-DM, SO-101, Trossen WidowX AI, Unitree G1, Universal Robots UR5e, UFACTORY xArm, XLeRobot.
Framework support
Checkpoints
- lerobot/eo1-base · LeRobot-org checkpoint; Hub license tag MIT.
- IPEC-COMMUNITY/EO-1-3B · Upstream release from the EO-1 project; Hub license tag MIT.
Related terms
Sources
- T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
LeRobot EO-1 docs: Qwen2.5-VL backbone plus continuous flow-matching action head; needs ≥1 camera, state and a language task.
“EO-1 uses a Qwen2.5-VL backbone for vision-language understanding and adds a continuous flow-matching action head for robot control.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/eo1.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
EO-1 data requirements list at least one image, observation.state, action and a language task.
“At least one visual observation, for example `observation.images.image`”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/eo1.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Policy type string is `eo1`.
“@PreTrainedConfig.register_subclass("eo1")”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/policies/eo1/configuration_eo1.py - T1 · upstream documentedrelease-notessource 2026-07-06verified 2026-10-24 (1d ago)lerobot 0.6.0
EO-1 was added in LeRobot v0.6.0 (absent from the v0.5.1 policies tree).
“* feat(policies): add EO-1 model by @HaomingSong in https://github.com/huggingface/lerobot/pull/3403”
https://github.com/huggingface/lerobot/releases/tag/v0.6.0 - T1 · upstream documentedpapersource 2025-08-28verified 2026-10-24 (1d ago)
EO-1 paper states the model comprises 3B parameters.
“a unified embodied foundation model comprising 3B parameters”
https://arxiv.org/abs/2508.21112v1 - T1 · upstream documentedofficial-docssource 2025-11-12verified 2026-10-24 (1d ago)
Upstream EO-1 README gives ~6.5GB GPU memory for inference with the pretrained model.
“you can load the model and run inference directly with just a few lines of code (requires ~6.5GB GPU memory)”
https://github.com/EO-Robotics/EO1/blob/451c363d7dff682700789afae0b9b17c004e8f8c/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 - T3 · community reportedhub-checkpointverified 2026-10-24 (1d ago)
Hub API: lerobot/eo1-base and IPEC-COMMUNITY/EO-1-3B exist, both tagged license:mit.
https://huggingface.co/lerobot/eo1-base