Learning Fault-Tolerant Locomotion with Adaptive Gait Timing

This paper presents a deep reinforcement learning controller for quadruped locomotion after sudden actuator power loss, targeting the limited actuation margins of heavier robots. The policy uses Proximal Policy Optimization with an asymmetric actor–critic architecture: during training, the critic receives privileged fault information, while the deployable actor reconstructs it from proprioceptive observation history through a latent-alignment Mean Squared Error loss. Its key innovation is adding a learnable gait-frequency action alongside joint-position targets, allowing step timing and contact scheduling to adapt to terrain and damage without predefined faulty-leg behaviors. Trained in MuJoCo XLA with 8192 parallel agents, the controller runs at 50 Hz on a 68 kg, 12-joint Kyon quadruped and learns to compensate for randomly timed single-joint torque failures on stepped terrain. Ablations show that increasing history from one observation to two provides the largest gain, while the selected H=3 offers a practical capacity–performance trade-off. In sim-to-sim tests, the policy generalized to previously unseen stairs with 10 cm steps and 0.7 m widths, plus ramps up to 13°, and it transferred zero-shot to the physical robot on flat ground. Training used an NVIDIA GeForce RTX 5090, while the reported hardware validation demonstrates stable locomotion under a sudden knee-joint power loss.