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

The gist: ReSYNC introduces Recovery-Driven Synthesis of Relational Concepts (ReSYNC), a method that progressively discovers and refines state abstractions from failure-recovery experience to support abstract planning, enabling robots to convert local recoveries into global failure avoidance at test time.

How it works

ReSYNC operates through an incremental dual-learning process that jointly learns skills and concepts from failures in training tasks The skill-learning phase involves the robot using reinforcement learning to learn to recover from failures seen in training tasks In the concept-learning phase, the robot discovers new relational predicates and refines its abstract planning model to explain and generalize the learned recovery behaviors This interaction allows ReSYNC to convert local recoveries seen during training into global failure avoidance at test time

Skill Learning

The process begins with Failure Mining, where for each training task Tj in T train, ReSYNC plans and executes with its current abstractions until failure, producing xFj To expand the few failures to a larger set, ReSYNC resamples oF relative to a relevant object oR derived from C F It then constructs a recovery learning MDP from DF and the original task goal g PPO is used to learn a recovery policy πCN and train a terminal function βCN via supervised learning Both the policy and the terminal function are relational, given a state x, ReSYNC applies a projection proj(x) containing only features of oR and oF The library is updated with C N, but the planner cannot yet leverage it without a symbolic description

Concept Learning

ReSYNC discovers concepts that support recovery skill reuse at test time after adding a new recovery skill C N to form C′ = C ∪ ∪C N This is achieved through Offline Concept Learning with IVNTR, which alternates between symbolic structure learning and neural classifier optimization To generate self-supervised data for concept learning, ReSYNC performs dreaming Dreaming creates counterfactual tasks by resampling oF relative to other objects This process allows the robot to collect data that captures skill reuse on other objects, such as generating data where the drawer is open but empty Concept learning involves learning predicates Ψ′ and operators OpC′ that enable abstract planning with C N

Progressive Learning and Generalization

ReSYNC enters later stages with the latest skills, concepts, and operators to progressively expand its library for increasingly complex unseen tasks Two designs support multi-stage concept learning and refinement: compositional dreaming and concept fine-tuning Compositional Dreaming involves composing failure objects oFv from previous stages with objects matching the types of oRv to produce increasingly diverse trajectories Concept Fine-tuning involves incrementally expanding the abstraction library using newly generated data while fine-tuning classifiers for existing concepts before new skills alter previously observed state distributions The system demonstrates robustness to curriculum ordering, as reversing the default task sequence yields performance comparable to the original ordering

Empirical Results

Across four physical domains, ReSYNC achieves an average success rate of 70% on compositionally novel tasks, outperforming strong baselines by over 50% In the Blocked Stacking domain, ReSYNC shows three strengths: (1) by leveraging learned relational abstractions, ReSYNC significantly outperforms baselines by 20 − 70% on T test i scenarios (2) ReSYNC is able to learn from multiple training stages without a significant performance drop as task horizon and variance grow (3) ReSYNC does not forget the previous tasks when progressively learning from new experiences In real-world domains, ReSYNC learns a non-prehensile recovery skill from failed picking attempts and discovers new predicates that ground this skill

Conclusion

ReSYNC represents a significant step toward robots that autonomously acquire abstractions for scalable, failure-aware planning in the physical world It substantially improves compositional generalization and sample efficiency over various prior baselines Furthermore, ReSYNC demonstrates sim-to-real transfer of ReSYNC on two long-horizon manipulation tasks where it successfully generalizes the usage of the learned non-prehensile recovery skill via discovered predicates and abstract planning

References

[1] K. Javed and R. S. Sutton. The Big World Hypothesis and Its Ramifications for Artificial Intelligence,page9

[2] S. Vats, M. Likhachev, and O. Kroemer. Efficient Recovery Learning using Model Predictive Meta-Reasoning,page3

[3] S. Vats, D. K. Jha, M. Likhachev, O. Kroemer, and D. Romeres. Recoverychaining: Learning Local Recovery Policies for Robust Manipulation.,page10

[4] A Bagaria, A. D. M Koch, R Rodriguez-Sanchez, S Lobel, and G Konidaris. Intrinsically Motivated Discovery of Temporally Abstract Graph-based Models of the World.,page2

[5] A Bagaria, A. D. M Koch, R Rodriguez-Sanchez, S Lobel, and G Konidaris. Intrinsically Motivated Discovery of Temporally Abstract Graph-based Models of the World.,page2

[6] A Li, N Kumar, T Lozano-Perez, and L P Kaelbling. Learning to Bridge the Gap: Efficient Novelty Recovery with Planning and Reinforcement Learning.,page3

[7] T. Silver, R Chitnis, N Kumar, W McClinton, T Lozano-Perez, L Kaelbling, and J B Tenenbaum. Predicate Invention for Bilevel Planning.,page11

[8] N Shah, J Nagpal, P Verma, and S Srivastava. From Reals to Logic and Back: Inventing Symbolic Vocabularies Actions and Models for Planning from Raw Data.,page23

[9] B Li, T Silver, S Scherer, and A Gray. Bilevel Learning for Bilevel Planning.,page10

[10] T Silver, A Athalye, J B Tenenbaum, T Lozano-Perez, and L P Kaelbling. Learning Neuro-Symbolic Skills for Bilevel Planning.,page12

[11] Y I Liu, B Li, B Eysenbach, and T Silver. SLAP: Shortcut Learning for Abstract Planning.,page12

[12] Y S Shao, Y Zheng, S Sun, P Chaudhari, V Kumar, and N Figueroa. Symskill: Symbol and Skill Co-Invention for Data-Efficient and Real-Time Long-Horizon Manipulation.,page13

[13] D Abel, D Arumugam, L Lehnert, and M Littman. State abstractions for lifelong reinforcement learning.,page22

[14] B Eysenbach, A Gupta, J Ibarz, and S Levine. Diversity is all you need: Learning skills without a reward function.,page11

[15] R K Nayyar and S Srivastava. Autonomous Option Invention for Continual Hierarchical Reinforcement Learning and Planning.,page22

[16] A Shukla, S Tao, and H Su. Maniskill-hab: A benchmark for low-level manipulation in home rearrangement tasks.,page10

[17] P W Battaglia, J B Hamrick, V Bapst, A Sanchez-Gonzalez, V Zambaldi, M Malinowski, A Tacchetti, D Raposo, A Santoro, R Faulkner. Relational Inductive Biases Deep Learning and Graph Networks.,page11

[18] A Vaswani et al. Attention is All You Need.,page3

[19] OpenAI. GPT-5.4 Thinking System Card, Mar 2026.,page7

[20] L P Kaelbling and T Lozano-Perez. Hierarchical Task and Motion Planning in the Now.,page11

[21] C R Garrett, T Lozano-Perez, and L P Kaelbling. Pddlstream: Integrating Symbolic Planners and Blackbox Samplers via Optimistic Adaptive Planning.,page12

[22] R Chitnis, T Silver, J B Tenenbaum, T Lozano-Perez, and L P Kaelbling. Learning Neuro-Symbolic Relational Transition Models for Bilevel Planning.,page11

[23] N Kumar, T Silver, W McClinton, L Zhao, S Proulx, T Lozano-Perez, L P Kaelbling, and J Barry. Practice Makes Perfect: Planning To Learn Skill Parameter Policies.,page12

[24] G Konidaris, L P Kaelbling, and T Lozano-Perez. From skills to symbols: Learning symbolic representations for abstract high-level planning.,page13

[25] Q Wang, B Li, Z Luo, Y Xu, A Gray, T Silver, S Scherer, K Sycara, and Y Xie.

1. The ReSYNC system can acquire the abstract knowledge needed to avoid them in the future by progressively discovering and refining state abstractions (relational predicates) from failure-recovery experience to support abstract planning. This allows it to convert local recoveries seen during training into global failure avoidance at test time.

2. The robot can generalize to a combinatorial number of unseen failures by recomposing the new skills and concepts that it has discovered, enabling it to solve long-horizon, previously unseen problems, as demonstrated by achieving success rates over 50% better than strong baselines in four simulated domains.

3. The system can perform sim-to-real transfer for non-prehensile manipulation skills and generalize to new scenarios through abstract planning, meaning it can apply learned skills in the physical world even when facing unseen scenarios.

4. The robot can utilize a skill-concept dual learning process where the skill-learning phase uses RL to recover from failures, and the concept-learning phase discovers new relational predicates that ground recovery skills in a new abstract planning model.

5. The system can achieve high planning efficiency by explicitly selecting predicates based on their impact on the objective, as ReSYNC learn[s] compact relational concepts that minimize the planning objective and yield single-plan execution, unlike DSG-M which relies solely on terminal predicates.

Intelligent robots should not only recover from failures, but also acquire the abstract knowledge needed to avoid them in the future. While reinforcement learning (RL) can learn reactive recovery behaviors, training a separate policy for every distinct failure mode is highly inefficient. We introduce Recovery-Driven Synthesis of Relational Concepts (ReSYNC), the first approach that progressively discovers and refines state abstractions (relational predicates) from failure-recovery experience to support abstract planning. Unlike purely reactive methods, ReSYNC jointly learns skills and concepts through an incremental dual-learning process. In the skill-learning phase, the robot uses RL to learn to recover from failures seen in training tasks. In the concept-learning phase, the robot discovers new relational predicates and refines its abstract planning model to explain and generalize the learned recovery behaviors. This interaction enables ReSYNC to convert local recoveries seen during training into global failure avoidance at test time. Across four simulated domains, we show that ReSYNC's ability to continually expand and refine its abstraction library allows it to solve long-horizon, previously unseen problems, outperforming strong baselines by over 50%. Additionally, we demonstrate sim-to-real transfer of ReSYNC, where it performs real-world non-prehensile manipulation skills and generalizes to unseen scenarios through abstract planning. Overall, ReSYNC represents a significant step toward robots that autonomously acquire abstractions for scalable, failure-aware planning in the physical world.