Learned robot policies often fail at high-precision, contact-rich manipulation, while classical motion planning generalises well but lacks contact awareness. This paper proposes a middle path: collect dense data autonomously in the narrow contact-rich part of a task, train local policies with offline reinforcement learning, and wrap them in a traditional motion planner for the full long-horizon operation.
On four real-world assembly tasks, the method needed only 2–2.5 hours of autonomous data collection per task. It achieved a 96% average success rate, beating a planning-only baseline (55%) and the strongest end-to-end learned baseline (37%). The local policies also reduced applied forces by 49%.
Notably, performance stayed at 90% in scene-level out-of-distribution tests, where end-to-end approaches struggle. The authors report more than 1,500 real-world evaluation rollouts in total, supporting the claim that targeted data collection can make contact-rich manipulation both reliable and sample-efficient.