Cerebellar-Inspired Residual Control for Fault Recovery: From Inference-Time Adaptation to Structural Consolidation

Robotic policies deployed in real-world environments often encounter post-training faults, where retraining, exploration, or system identification are impractical. We introduce an inference-time, cerebellar-inspired residual control framework that augments a frozen reinforcement learning policy with online corrective actions, enabling fault recovery without modifying base policy parameters. The framework instantiates core cerebellar principles, including high-dimensional pattern separation via fixed feature expansion, parallel microzone-style residual pathways, and local error-driven plasticity with excitatory and inhibitory eligibility traces operating at distinct time scales. These mechanisms enable fast, localized correction under post-training disturbances while avoiding destabilizing global policy updates. A conservative, performance-driven meta-adaptation regulates residual gain and plasticity, preserving nominal behavior and suppressing unnecessary intervention. Experiments on MuJoCo benchmarks under actuator, dynamic, and environmental perturbations show improvements of up to $+66$% on $\texttt{HalfCheetah-v5}$ and $+53$% on $\texttt{Humanoid-v5}$ under moderate faults, with graceful degradation under severe shifts and complementary robustness from consolidating persistent residual corrections into policy parameters.

Robots trained in simulation or controlled laboratory settings can fail when they are deployed in new conditions, such as when an actuator becomes weaker, the ground becomes slippery, or the robot’s dynamics change unexpectedly. In many real-world settings, it is not practical or safe to retrain the robot, let it explore new actions, or identify the exact fault during deployment. This paper introduces a lightweight adaptation method inspired by the cerebellum, the part of the brain involved in fast motor correction. The method keeps the robot’s original learned controller fixed and adds a small corrective controller that only becomes active when performance drops. This correction is bounded, local, and designed to avoid interfering with normal behavior. Across simulated locomotion and manipulation tasks, the method helps robots recover from unexpected faults more effectively than several robustness and adaptation baselines. The results suggest that fast, conservative correction can make learned robot controllers more reliable under deployment-time changes without requiring retraining.