Learning Where it Matters: Autonomous Data Collection and Offline Reinforcement Learning for Local Contact-Rich Manipulation
End-to-end robot policies often fail in high-precision tasks due to brittle data collection and poor generalization. This work proposes a modular approach: autonomously collecting dense data specifically in the critical, contact-rich segments of a task, then using offline reinforcement learning to create robust local policies. Integrating these with traditional motion planning achieves high success rates for long-horizon assembly.