Terrain-Height-Blind Quadrupedal Locomotion Through Privileged Teacher–Student Learning and a Shared Gait-Structured Central Pattern Generator

Terrain-aware training can support locomotion without requiring terrain-height input at deployment. We present a quadrupedal control framework that combines privileged teacher–student learning with a shared gait-structured sensor-consistency central pattern generator (GSC-CPG). A recurrent student uses proprioception, contact, commands, action history, and rhythmic state to produce 12-dimensional modulation of the controller. Teacher and student act through the same phase and foot-target pathways. We evaluated the framework in Go2 simulations with disturbances applied to actor observations and direct simulator feedback retained downstream. Across three matched policy pairs, each with a shared behavior-cloned initialization, continued teacher guidance reduced the pooled rough-terrain failure rate from 1.479 to 1.097 events per robot-minute. Tracking, body tilt, forward speed, and action saturation also improved, while action variation increased. These locomotion benefits extended to a held-out roughness amplitude. On an unseen 5° uphill slope, guided policies had no physical falls over 384 robot-minutes, compared with nine for guidance-off policies. Component tests showed selective, reversible phase responses and a support-related speed–stability trade-off. The results support a shared rhythmic interface for transferring terrain-aware supervision to terrain-height-blind locomotion under the tested simulation conditions.