Does ACT run on SO-100?
works
Capabilities
| Capability | State | Notes | Evidence |
|---|---|---|---|
| Train / fine-tune | works | ev-smolvla-paper-so100 | |
| Real-robot inference | works | ev-smolvla-paper-so100 | |
| Simulation inference | no verified evidence | — | |
| Teleop + record | works | ev-smolvla-paper-datasets |
Evidence 2 sources; the cell shows the highest tier
- T1 · upstream documentedpapersource 2025-06-02verified 2026-10-24 (1d ago)
SmolVLA paper (Hugging Face/LeRobot team), Table 3: real-world SO100 benchmark (pick-place, stacking, sorting) comparing ACT (single-task), pi0 3.5B and SmolVLA (multi-task).
“Table 3: Real-world benchmarks (SO100). Success rate (%) across three tasks using policies trained in multi-task and single-task settings.”
https://arxiv.org/html/2506.01844v1 - T1 · upstream documentedpapersource 2025-06-02verified 2026-10-24 (1d ago)
SmolVLA paper: real-world evaluation datasets were recorded on SO-100 (three tasks) and SO-101 (one task).
“For real-world evaluation, we collected three datasets using the SO-100 robot arm”
https://arxiv.org/html/2506.01844v1
Embed this cell
Badge for READMEs and model cards. It always shows the current state.
[](https://caniembody.pages.dev/compat/so100/act)<iframe src="https://caniembody.pages.dev/embed/so100/act" width="380" height="132" style="border:0" title="ACT on SO-100 | CanIEmbody"></iframe>History
- 2026-10-25 fecccd1 Initial release: robot policy compatibility matrix · Danilo Znamerovszkij