ReSYNC: New Approach Enables Robots to Learn from Failures and Avoid Future Errors

Researchers have introduced Recovery-Driven Synthesis of Relational Concepts (ReSYNC), a novel approach that allows robots to not only recover from failures but also develop abstract knowledge to prevent similar failures in the future. The method addresses a key limitation in current robotics: training separate policies for each distinct failure mode is inefficient and does not scale well.

ReSYNC employs a dual-learning process that integrates skill acquisition with concept discovery. During the skill-learning phase, robots use reinforcement learning to recover from failures observed during training. In the concept-learning phase, robots discover new relational predicates and refine their abstract planning models to explain and generalize these recovery behaviors. This integrated approach converts localized recoveries during training into global failure avoidance during testing.

Evaluation across four simulated domains demonstrates that ReSYNC's ability to continually expand and refine its abstraction library enables it to solve long-horizon problems never encountered before, outperforming strong baseline approaches by over 50%. The researchers also demonstrated successful sim-to-real transfer, where the system performed real-world non-prehensile manipulation tasks and generalized to previously unseen scenarios through abstract planning.

The research represents a significant advancement toward creating robots that can autonomously acquire abstractions for scalable, failure-aware planning in physical environments. By learning both reactive skills and abstract planning concepts from failure experiences, ReSYNC enables more efficient and generalizable robot learning that scales beyond individual failure scenarios.