Systematic emergence of recovery behaviors in VLA models

Determine whether recovery behaviors systematically emerge in most deployment settings for generalist Vision-Language-Action (VLA) robot control models, and establish a rigorous evaluation procedure—such as plotting test-time scaling curves of recovery maneuvers—to assess the presence and consistency of this emergence across tasks and environments.

Background

RaC emphasizes training robot policies on interventions that pair recovery with correction, showing that such behaviors lead to robustness and test-time scaling on long-horizon tasks. The authors propose exploring how these ideas relate to generalist Vision-Language-Action (VLA) models, which are trained across diverse tasks and embodiments.

While prior results indicate isolated instances of recovery behavior in VLA models, it is unknown whether this behavior arises consistently and systematically across typical deployment scenarios. Clarifying this would inform the community on the reliability of VLA models in handling out-of-distribution states and compounding errors, and would guide evaluation methodologies that capture test-time scaling via retries and recovery maneuvers.

References

Finally, while prior results do show some examples of recovery behaviors in VLA models, it is unclear if such behaviors systematically emerge in most settings or not, and studying this aspect rigorously (for example, by plotting test-time scaling curves analogous to Figure~\ref{fig:lid_retries_vs_success}) is also useful for the community.

— RaC: Robot Learning for Long-Horizon Tasks by Scaling Recovery and Correction  (2509.07953 - Hu et al., 9 Sep 2025) in Section 6: Discussion, Conclusion, and Future Work (Future work paragraph)

Whether more targeted interventions can turn this attention shift into reliable recovery remains an open question.

— Blackout vs. Freeze: Analyzing Physical Failure Modes of VLAs under Camera Faults  (2609.39145 - Suh et al., 30 Sep 2026) in Section 5.1, subsection “Training-Based Mitigation: Camera Blackout Training”

This test also increases the corpus from roughly $450$ to $600$ images and adds another layer of distractors. It measures transfer to longer, harder searches, rather than isolating hop count alone. Training on four-hop worlds is a possible next experiment; recovery from such training has not been established here.

— Learning to Route in Visual Space via Multi-Step Embedding Retrieval  (2609.38743 - Chen et al., 30 Sep 2026) in Section 'Generalization and Limitations', subsection 'Four-Hop Extrapolation' (Appendix, Section 'Generalization and Limitations', subsection 'Four-Hop Extrapolation')

Household and production work is longer and less forgiving: tens of operations over a changing world state, objects found rather than placed, and success measured in throughput and rework rather than a binary predicate. Our failure analysis indicates what such a suite must measure beyond success, namely whether a policy tracks which steps are complete, notices that an earlier step went wrong, and can recover without restarting the episode.

— HandsOn: A Large-Scale Simulation Benchmark for Generalizable Dexterous Manipulation  (2610.26765 - Luo et al., 19 Oct 2026) in Section 6, Future Work, paragraph “Long-horizon household and production tasks”