ReSYNC: Robots learn abstract concepts from their own failures

Intelligent robots should not just recover from failures but also learn why they happened and how to avoid them in the future. Traditional reinforcement learning (RL) trains separate policies for each failure mode, which is inefficient and doesn't generalize. Researchers from Carnegie Mellon University, MIT, and other institutions present ReSYNC (Recovery-Driven Synthesis of Relational Concepts), the first framework that jointly learns skills and abstractions from failure-recovery experience. In the skill-learning phase, the robot uses RL to learn reactive recovery behaviors for observed failures. In the concept-learning phase, it discovers new relational predicates—abstract state features—that explain and generalize those recoveries. This incremental dual-learning process builds a growing library of skills and abstract planning models, allowing the robot to convert local recoveries into global failure avoidance at test time.

Across four simulated domains, ReSYNC consistently outperformed strong baselines (including pure RL and hierarchical RL) by over 50% on long-horizon, previously unseen problems. The key advantage is its ability to continually expand and refine its abstraction library as it encounters new failures. Crucially, the team demonstrated sim-to-real transfer: the same system controlled a real robot arm performing non-prehensile manipulation tasks like pushing and sliding objects to targets, generalizing to novel scenarios through abstract planning without any real-world fine-tuning. ReSYNC represents a significant step toward robots that autonomously acquire the conceptual understanding needed for scalable, failure-aware planning in the physical world.