VPIES: Variational Privileged Information Encoder as Scaffold for Legged Locomotion Learning

Legged robots face significant challenges in complex terrains due to partial observability. While teacher-student frameworks address this through imitation, they often cause representation mismatch and covariate shift, limiting deployment robustness. To address these limitations, we propose the Variational Privileged Information Encoder as Scaffold (VPIES), a single-stage reinforcement learning framework that consists of a variational encoder and a dual-timescale proprioceptive history encoder. VPIES encodes privileged states into a variational latent for guided exploration, with adaptive KL-annealing that progressively removes this scaffold for privilege-free deployment. To support the resulting handover and capture multi-scale dynamics, VPIES adopts a dual-timescale proprioceptive encoder that combines short-horizon reactive control with long-horizon phase-consistent adaptation from proprioception alone, thereby eliminating the need for explicit environment identification. In simulation, VPIES reduces velocity tracking error and improves out-of-distribution robustness by approximately 20% compared to strong baselines. Real-world experiments on the Unitree Go2 achieve higher terrain success rates, and deployment on the humanoid G1 demonstrates stable and agile locomotion, including stair climbing, confirming the practicality and robustness of VPIES for legged systems.