· Personal experiment
Getting ACT running on a real arm
I trained ACT on nine demonstrations I collected by teleoperating my SO‑101 arm. In this early experiment, the policy picks up an eraser and places it into a box.
Setup
- Data
- 10 teleoperated demonstrations, without the final return-to-home motion. I excluded one with an anomalous start and trained on the remaining 9.
- Training
- ACT, 20,000 training steps.
- Inputs
- One image from each of two cameras and a 6-dimensional joint state, predicting the next 30 actions.
- Execution
- Execute all 30 actions before predicting again; no temporal ensembling; a 15-second rollout budget. Small predicted excursions beyond joint limits were clipped to the valid range.
A deployment observation
Increasing the shoulder and elbow target-offset limits from 2° to 5° made lifting smoother in the later run, with the model and training data unchanged. These offsets are relative to the current joint angles; the other arm joints remained limited to 2°.
Same policy and training data · 1× speed
Both clips are shown at their recorded speed, with lead-in and trailing footage trimmed and audio removed.