Privileged Critic Training Enables Sensor-Free Thruster Fault Adaptation in End-to-End RL
2026-08-24 • Robotics
Robotics
AI summaryⓘ
The authors studied how to help robots that use thrusters navigate even when some thrusters fail or get stuck. Instead of needing extra sensors to detect these problems, they trained a system where the 'critic' part knows the true state of thruster failures during training, while the 'actor' part only sees normal sensor info. This approach, called RAFT, uses memory and learns to adapt to failures without special sensors at run-time. When tested on a robot with several thruster problems, RAFT performed much better than systems that ignore failures and almost as well as ones that have full failure information.
fault-tolerant navigationthruster failuresreinforcement learningPPO (Proximal Policy Optimization)privileged critic trainingrecurrent policiesrobot controlasymmetric actor-criticfault adaptationfloating platform robot
Authors
Ricard Marsal I Castan, Miguel A. Olivares-Méndez
Abstract
Fault-tolerant navigation for thruster-actuated robots requires online adaptation to failures that are neither binary nor fully observable: thrusters may degrade continuously, fail dead, or jam stuck-open. Classical fault detection pipelines require dedicated sensors unavailable at deployment; oracle controllers that observe the true failure state are equally impractical. We show that privileged critic training is sufficient for sensor-free fault adaptation: giving the PPO value function access to the true degradation state dgt during training, while the actor receives only standard task observations, shapes a policy that compensates for failures at deployment without any dedicated fault sensing. We propose RAFT (Recurrent Asymmetric Fault Tolerant), a policy with recurrent memory trained with a privileged asymmetric critic. Evaluated on a floating-platform robot (8 thrusters, 1 reaction wheel) under up to four simultaneous thruster failures, RAFT achieves 70.2% success at four concurrent failures, closing 84% of the gap from a failure-naive baseline (4.8%) to an oracle policy that sees the full degradation state at deployment (82.4%). All code, checkpoints, and data are open-source.