All-Terrain Quadrupeds: A Unified Policy for Locomotion, Fall & Push Recovery with Latent Estimation

This paper presents a single-policy reinforcement-learning framework that enables a quadruped robot to walk, run, climb irregular stairs, traverse rough terrain, and perform agile self-righting after falls or strong pushes - entirely without exteroceptive vision or hand-coded mode switches. The policy is trained end-to-end in simulation together with a contrastive latent-information encoder that extracts terrain and disturbance cues directly from short proprioceptive histories. A curriculum and broad domain randomization yield a controller that transfers to hardware with no additional tuning. Extensive tests in both the simulator and on a real Unitree A1 confirm that the same policy performs all target behaviors seamlessly across previously unseen environments, demonstrating that proprioception-only learning can produce truly self-reliant quadrupeds with minimal engineering effort.