SciGroveBeta
Robotics

Recover, Discover, Plan: Learning Skills and Concepts from Robot Failures

Bowen Li, Mayank Mishra, Y. Isabel Liu, Stone Tao, Nishanth Kumar, Alexander G. Gray, Ruwan Wickramarachchi, Jonathan Francis, Sebastian Scherer, Tom Silver

Featured June 19, 2026

Original Paper

This analysis was generated by SciGrove. Upload your own PDFs or enter a DOI — and get the same AI breakdown on any paper.

Get started

AI-generated analysis — This is SciGrove's AI interpretation of the paper, not peer-reviewed content. Always refer to the original paper.

Simply

Robots learn to fix their own mistakes by turning every failure into a new, reusable skill and a smart rule for future planning.

In depth
ReSYNC introduces a dual-learning process where robots treat failures as data for discovery. By learning recovery skills via reinforcement learning and simultaneously synthesizing relational predicates through self-supervised dreaming, the system enables robots to transition from reactive recovery to proactive, abstract planning.

Key Takeaways

  • 1
    ReSYNC converts local failure-recovery experiences into global failure avoidance by updating the robot's symbolic planning model.
  • 2
    The compositional dreaming process generates diverse, self-supervised data to refine abstract concepts without requiring human demonstrations.
  • 3
    The framework achieves significant sample efficiency and generalization, outperforming baselines by over 50% in long-horizon, unseen tasks.

Conceptual Flow

HIGH LEVEL
1
Methodology: The Dual-Learning Loop

The robot learns a new trick when it fails, then figures out a rule to explain why that trick works.

Failure Experience
Learn Skill and Concept
Updated Planning Model
2
Results: Impact on Performance

Robots that learn from mistakes solve much harder puzzles than robots that just try to fix things one by one.

Reactive Robot
Outperforms
ReSYNC Robot