Researchers led by Zhaoting Li, with Cosimo Della Santina and Jens Kober, have built SeRA, a way for robots that learn from human corrections during use to also learn from their own autonomous attempts, according to a paper on arXiv. Such attempts are plentiful but have no human labels, and copying the robot's past actions can reinforce mistakes.

Later versions of the policy relabel the earlier experience, relabels are kept only when several versions agree, and because agreement does not prove they are right, SeRA builds a set of acceptable actions around each one. A diffusion policy is then trained to make actions inside those sets more likely, combining them with the available human corrections.

In five simulated manipulation tasks it beat the Set-Supervised Diffusion Policy it builds on and methods that imitate past robot actions. In a real tomato-picking task it succeeded 80% of the time, against 40% for that policy, with the same human feedback and training budget.

Source: arXiv Robotics (cs.RO)

Background

Zhaoting Li

Robotics researcher

Robotics researcher.