Reproduce “Learning Multi-Stage Tasks with One Demonstration via Self-Replay” on OpenResearch

About this paper

Norman Di Palo, Edward Johns

Researchers at Imperial College London developed Self-Replay, a robot learning method that enables manipulators to learn complex, multi-stage tasks from a single human demonstration. It leverages autonomous data collection to train specialized neural networks, achieving 70-80% success rates across various tasks while drastically reducing human effort compared to existing imitation learning techniques.

Key insights

• Complex multi-stage manipulation tasks can be effectively decomposed into alternating phases of learned object-reaching ('bottleneck pose') and open-loop action replay for interaction.
• Autonomous, self-supervised data collection through 'Self-Replay' drastically reduces human effort and data requirements, making robot skill acquisition more accessible.
• Active sampling strategies (Active Self-Replay) enhance learning efficiency and robustness by prioritizing data from challenging or under-explored workspace regions.

Summary by alphaXiv.