The paper introduces Action Chunking with Transformers (ACT), an imitation learning method that learns a generative model over short sequences of actions (chunks), trained from teleoperated demonstrations on a low-cost bimanual rig.
Findings
- Six real-world fine manipulation tasks, such as opening a translucent condiment cup and slotting a battery.
- 80 to 90% success with only 10 minutes of demonstrations per task.
Limits
- Abstract-level reading only.
Relevance
ACT is a standard baseline in later benchmarks; in RoboTwin 2.0 it averages 29.7% on clean tasks and 1.7% under randomisation. Executing a chunk commits the robot to several steps at once, so an open question for failure-recovery is how quickly a chunked policy can react when something breaks mid-chunk. Available in LeRobot.
Links
- arXiv
- Subfield: Robot learning