MoE-DAgger picks when to ask a human to correct a robot during multi-phase tasks

Researchers led by Sungjun Heo have presented MoE-DAgger, a rule for deciding when a person should step in to correct a robot learning manipulation tasks with ordered phases, in a paper posted on arXiv.

It relies on the router of a mixture-of-experts policy. High router entropy can mean a normal handoff between adjacent phases or a policy falling back to an earlier phase, the authors write, so MoE-DAgger looks at the router's recent history to tell them apart, without giving the policy a phase label. It lets normal transitions settle and asks for a correction when high entropy persists, a transition times out or routing keeps favouring an earlier phase.

With matched human-control budgets, it reached 84.0% task success on the Green task and 78.0% on Red, 22 and 26 percentage points above entropy filtering that ignores phases, with no collisions seen in 50 rollouts per task.