Physical AI Lab // Episode 1.6 // One-pager
Imitation learning: Learning from demos
Just copy it? Why it's not that simple
Three key points
- Imitation learning means copying demonstrations (behaviour cloning): learning "what you see, what you do". But once the robot drifts, it enters unfamiliar states and the errors build up and snowball (compounding error).
- Two well-known fixes: action chunking (ACT learned 6 real tasks with about 10 minutes of demonstrations; 80–90% success on tasks such as opening a condiment cup and slotting a battery, R) predicts a short stretch at once and is steadier; Diffusion Policy (about 46.9% average improvement over the previous best methods across 12 tasks, R) handles "several correct ways to act in the same scene".
- More demonstrations aren't always better: a data scaling study found diversity matters more than the absolute number; with enough diverse data, two tasks reach about 90% success in new environments with unseen objects (R). The curve rises steeply, then flattens.
Evidence: O official · R research · M media
| R | 80–90% | success rate of ACT (action chunking) on real tasks such as opening a condiment cup and slotting a battery, with only about 10 minutes of demonstrations (paper) |
| R | 46.9% | average improvement of Diffusion Policy over the previous best methods, across 12 tasks from 4 benchmarks (paper) |
| R | About 90% | success rate for two tasks in new environments with unseen objects, once enough diverse data was collected; the paper says diversity matters more than the absolute number of demonstrations (paper) |
Further reading
- ACT / ALOHA: low-cost bimanual imitation learning (arXiv 2304.13705)
https://arxiv.org/abs/2304.13705 - Diffusion Policy (arXiv 2303.04137)
https://arxiv.org/abs/2303.04137 - Data Scaling Laws in Imitation Learning (arXiv 2410.18647)
https://arxiv.org/abs/2410.18647