Event
TIDE: A Human-in-the-Loop Online Post-Training System for Robot Manipulation
1 reports · 1 sources
Chun Gu and other researchers propose TIDE (Training In DEployment), a human-in-the-loop online post-training system. It adopts a source-conditioned supervised learning framework, using human interventions and policy rollouts during deployment to update a robot's dual-arm manipulation policy online, without needing to learn a critic, reward model, or action scorer. On four real dual-arm manipulation tasks, TIDE outperforms both the warm-start policy and matched additional offline demonstrations on milestone-normalized task progress and binary full-task completion rate. Each task-specific online post-training session, from the start of deployment rollouts to the final checkpoint being available, takes less than one hour, with learning and checkpoint processing proceeding asynchronously.
Report timeline
- TIDE: An online post-training system for human-in-the-loop deployment in robot manipulation
The authors propose TIDE (Training In DEployment), a human-in-the-loop online post-training system that uses a provenance-conditioned supervised learning framework to update the policy online from human interventions and policy rollouts during deployment, without learning a critic, reward model, or action scorer.
arXiv cs.RO ·