Does ACT run on Reachy 2?
Capabilities
| Capability | State | Notes | Evidence |
|---|---|---|---|
| Train / fine-tune | works | ev-lerobot-reachy2-train, ev-hub-erlhub-act-reachy2 | |
| Real-robot inference | works with caveats | Community deployment via lerobot-record --policy.path on a physical Reachy 2; scored 5 on an undefined /10 scale. | ev-reachy-pollen-act-deploy, ev-reachy-pollen-registry |
| Simulation inference | no verified evidence | — | |
| Teleop + record | works | Recording uses --robot.type=reachy2 with --teleop.type=reachy2_teleoperator; the page also documents Pollen's own VR data-acquisition app. | ev-lerobot-reachy2-record |
Caveats
- On-robot inference is community-reported only. Upstream documents training and recording; its evaluation step uses lerobot-eval, which in v0.6.1 evaluates in a simulation env, so it does not document on-robot inference.
- The only result is a score of 5 under a 'Performance (10/10)' header in a community README, with no stated protocol; the README does not say explicitly that the score came from physical trials, though its documented ACT deployment targets the physical robot r2-0008.
- Tested configuration: both arms, neck and torso camera at 640x480; mobile base, antennas and teleop cameras disabled. The LeRobot version used is not stated.
Evidence 6 sources; the cell shows the highest tier
- T3 · community reportedhub-checkpointsource 2026-04-13verified 2026-10-25 (today)
config.json type=act; train_config.json dataset erl-hub/reachy-pick-and-place-images whose meta/info.json robot_type=reachy2.
https://huggingface.co/erl-hub/reachy-act-execute-pick-and-place - T3 · community reportedsource-codesource 2026-04-14verified 2026-10-25 (today)
Community README's 'Deploying policies on reachy2' section runs lerobot-record with --robot.type=reachy2 on robot r2-0008 at 192.168.10.172 (the same address its VR-teleoperation section uses for the physical robot) and --policy.path=erl-hub/reachy-act-execute-pick-and-place.
“--policy.path=erl-hub/reachy-act-execute-pick-and-place”
https://github.com/NikolaRaicevic2001/reachy_pollen/blob/ae76a3eedbe07a7a900262b31e986e2e4bec45a2/README.md - T3 · community reportedsource-codesource 2026-08-21verified 2026-10-25 (today)
The README's policy registry ('policies you train or deploy') scores the ACT run trained on erl-hub/reachy-pick-and-place-images 5 under a 'Performance (10/10)' header; no scoring protocol is given.
“| `act` | Action Chunking Transformer (`--policy.type=act`) | 250MB | 2026-04-14_01-02-26-erl-hub-reachy-pick-and-place-images_s7908 | CPU; Intel Core 9 (RTX GeForce 2080+) | 5 |”
https://github.com/NikolaRaicevic2001/reachy_pollen/blob/7edb3a080d1cd68d88a0f03a518589260e9110dd/README.md - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
The Reachy 2 page's 'Step 3: Train' runs lerobot-train with --policy.type=act on the recorded Reachy 2 dataset.
“--policy.type=act \ --output_dir=outputs/train/reachy2_test”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/reachy2.mdx - T1 · upstream documentedofficial-docssource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
Upstream states Reachy 2 is supported by LeRobot's recording features.
“Reachy 2 is fully supported by LeRobot’s recording features.”
https://github.com/huggingface/lerobot/blob/v0.6.1/docs/source/reachy2.mdx - T1 · upstream documentedsource-codesource 2026-08-03verified 2026-10-24 (1d ago)lerobot 0.6.1
In v0.6.1, EvalPipelineConfig (the lerobot-eval CLI config) has a required env field and no robot field.
“env: envs.EnvConfig eval: EvalConfig = field(default_factory=EvalConfig) policy: PreTrainedConfig | None = None”
https://github.com/huggingface/lerobot/blob/v0.6.1/src/lerobot/configs/eval.py
Embed this cell
Badge for READMEs and model cards. It always shows the current state.
[](https://caniembody.pages.dev/compat/reachy2/act)<iframe src="https://caniembody.pages.dev/embed/reachy2/act" width="380" height="132" style="border:0" title="ACT on Reachy 2 | CanIEmbody"></iframe>