DemoSpeedup: Speeding Up Robot Learning 3x by Dropping Unimportant Frames

Forum topic · 二一 · 2026-05-13

Summary

A CoRL 2025 Oral paper called DemoSpeedup addresses the problem of robots learning overly slow policies from human demonstrations. Human demos are cautious and slow, so imitation-learned robots inherit slow-motion strategies. DemoSpeedup measures the 'action entropy' of every demonstration frame: frames with low entropy (the model is confident across action options) are precision-critical keyframes, while high-entropy frames indicate casual, skippable motions. By pruning high-entropy frames, the method compresses demonstration data and retrains the policy, achieving up to 3x faster robot execution. Notably, task success rates improve rather than drop, attributed to shorter decision sequences reducing cumulative error. The key insight is not simply 'acting faster' but knowing when to slow down.

Robots learn manipulation skills by imitating human demonstrations, but human demos tend to be slow — cautious and frame-by-frame. As a result, the robot learns a slow-motion policy as well.

The CoRL 2025 Oral paper DemoSpeedup offers an elegant solution: measure the "action entropy" of each frame — if the model is highly confident about all action options at a given frame (low entropy), it is a precision-critical "keyframe"; if the probability distribution is spread out (high entropy), the action at that frame is relatively casual and can be sped up.

This insight — that high-entropy frames can be safely "fast-forwarded" — is used to compress the demonstration data. Low-entropy frames are kept; high-entropy frames are skipped. Retraining on the compressed data makes the robot's execution up to 3x faster, and task success rates don't drop — they actually improve, because shorter decision sequences reduce accumulated error.

*Feynman-style takeaway*: the key isn't "doing things faster" — it's "knowing when to slow down."

Reference: DemoSpeedup / CoRL 2025 Oral / arXiv:2605.XXXXX

DemoSpeedup: Speeding Up Robot Learning 3x by Dropping Unimportant Frames

Summary

Tags