Does ACT run on Franka Research 3?
Capabilities
| Capability | State | Notes | Evidence |
|---|---|---|---|
| Train / fine-tune | works | ACT fine-tuned on PhAIL's 449-episode FR3 teleoperation corpus with its default open-source recipe; the checkpoint path sits under phail_unified/lerobot/. | ev-phail-card, ev-phail-act-sample-timeout, ev-hub-sashopepi-act |
| Real-robot inference | works with caveats | ev-phail-card, ev-phail-act-sample-timeout, ev-phail-act-sample-success | |
| Simulation inference | no verified evidence | — | |
| Teleop + record | no verified evidence | — |
Caveats
- Community-reported only (anonymous PhAIL benchmark under double-blind review); not documented upstream or maintainer-tested.
- Ran through the benchmark authors' own inference server (Positronic stack) with end-effector-frame actions at 15 Hz, not through a native LeRobot FR3 robot type.
- The dataset card states no per-policy success rates; of the two ACT rollouts in its sample subset, one moved all 8 items and one ran out of time at 240 s with 5 of 8 moved.
Evidence 4 sources; the cell shows the highest tier
- T3 · community reportedcommunity-postsource 2026-05-06verified 2026-10-25 (today)
PhAIL v1.0 dataset card: ~950 real-robot episodes on a Franka Research 3 with Robotiq 2F-85 (bin-to-bin pick-and-place), including ~150 ACT evaluation rollouts using ACT's default fine-tuning recipe.
“Action Chunking Transformer | `inference` | ~150 | Default fine-tuning recipe”
https://huggingface.co/datasets/phail-anon/phail-v1.0/blob/5bc590213c5d2615eefeb47bc23a700f8c57022a/README.md - T3 · community reportedcommunity-postsource 2026-05-06verified 2026-10-25 (today)
Sample ACT rollout metadata (model=act, server type act, checkpoint .../phail_unified/lerobot/270226-ee/checkpoints/220000/pretrained_model, URDF of an FR3): Batteries, 5 of 8 items moved, outcome Ran out of time after 240 s.
“"eval.outcome": "Ran out of time"”
https://huggingface.co/datasets/phail-anon/phail-v1.0/blob/5bc590213c5d2615eefeb47bc23a700f8c57022a/sample/inference/000000000000/000000000003/static.json - T3 · community reportedcommunity-postsource 2026-05-06verified 2026-10-25 (today)
Sample ACT rollout metadata (same checkpoint): Wooden spoons, 8 of 8 items moved in ~199 s, outcome Success.
“"eval.outcome": "Success"”
https://huggingface.co/datasets/phail-anon/phail-v1.0/blob/5bc590213c5d2615eefeb47bc23a700f8c57022a/sample/inference/000000000000/000000000061/static.json - T3 · community reportedhub-checkpointsource 2026-03-27verified 2026-10-25 (today)
config.json type=act; train_config.json dataset SashoPepi/tomato_close_box (20000 steps), whose meta/info.json robot_type=fr3 (257 episodes, 8-D state, two cameras). Auto-generated card, no results.
https://huggingface.co/SashoPepi/tomato_close_box_20Ksteps
Embed this cell
Badge for READMEs and model cards. It always shows the current state.
[](https://caniembody.pages.dev/compat/franka-fr3/act)<iframe src="https://caniembody.pages.dev/embed/franka-fr3/act" width="380" height="132" style="border:0" title="ACT on Franka Research 3 | CanIEmbody"></iframe>History
- 2026-10-25 fecccd1 Initial release: robot policy compatibility matrix · Danilo Znamerovszkij