Learning-based fault-tolerant control for flexible joint robotic manipulators: An actor-critic framework with experimental verification.

Flexible-joint robotic manipulators (FJRMs) offer significant advantages over traditional rigid-link counterparts, including improved dexterity, greater motion flexibility, and lower energy consumption. However, their control remains challenging due to inherent nonlinear dynamics, elastic joint coupling, external disturbances, and susceptibility to actuator faults. This paper presents a Lyapunov-guided actor-critic reinforcement learning (RL) fault-tolerant control strategy for FJRMs. Different from conventional adaptive neural-network control, where the neural network is mainly used as a direct uncertainty approximator, the proposed method introduces a critic network to learn the long-run cost-to-go from tracking error, control effort, and fault-induced performance degradation, and uses this critic evaluation to guide actor policy improvement online. The Lyapunov design supplies a stabilizing control structure, while the actor-critic learning mechanism provides value-based policy adaptation under unknown matched uncertainties and actuator faults. Semi-global uniform ultimate boundedness of all closed-loop signals is established, and experiments on a Baxter robotic platform demonstrate improved tracking accuracy and robustness compared with PID and single neural-network controllers under faulty conditions.