Research paper

Recovery RL · Separate task and recovery policies

Offline constraint-violation data trains a recovery mechanism that can intervene in a task policy. Evaluation combines six simulated tasks with image-based obstacle avoidance on a physical robot.

Reviewed 2026-11-05 · Source and bibliography checked

Bibliographic details and the editorial overview were checked against the original source. Access conditions and editions are identified with the source.

Original title: Recovery RL: Safe Reinforcement Learning With Learned Recovery Zones
Publisher: Institute of Electrical and Electronics Engineers (IEEE)
Authors: Brijen Thananjeyan, Ashwin Balakrishna, Suraj Nair, Michael Luo, Krishnan Srinivasan, Minho Hwang, Joseph E. Gonzalez, Julian Ibarz, Chelsea Finn, Ken Goldberg
DOI: 10.1109/lra.2021.3070252
First online publication: 2021-03-31
Source language: en
Source country: US
Edition: IEEE Robotics and Automation Letters
Access: Publisher metadata and open author manuscript
Publication status: Published

Source overview

Offline constraint-violation data trains a recovery mechanism that can intervene in a task policy. Evaluation combines six simulated tasks with image-based obstacle avoidance on a physical robot.

Use and review

Empirical task/constraint results are not a no-injury guarantee. Data coverage, recovery reliability, sensing and execution limits need application-specific validation.

Sources & editions

Recovery RL: Safe Reinforcement Learning With Learned Recovery Zones ↗

Institute of Electrical and Electronics Engineers (IEEE) · IEEE Robotics and Automation Letters

Archiveabstract/authors and IEEEprimaryindexedabstract checked; IEEE states firstpublication2021-03-31, July2021issue. Empirical reductions in constraintviolations are not zero-risk guarantees.