RPL: Learning Robust Humanoid Perceptive Locomotion on Challenging Terrains

TL;DR

RPL tackles robust, multi-directional humanoid perceptive locomotion on challenging terrains with payloads by separating learning into terrain-specific experts and a unified depth-based transformer policy. Stage 1 leverages privileged height maps to master decoupled locomotion and loco-manipulation under end-effector perturbations, while Stage 2 distills into a multi-view depth policy that fuses inputs from multiple cameras. Key innovations include an efficient multi-depth rendering system that speeds up depth synthesis by about 5×, and robustness techniques—Depth Feature Scaling Based on Velocity Commands (DFSV) and Random Side Masking (RSM)—that improve performance under asymmetric observations and unseen terrain widths. Extensive simulation and real-world experiments on a Unitree G1 demonstrate stable long-horizon bidirectional locomotion with a 2 kg payload across slopes, stairs, and stepping stones, validating RPL’s practicality and robustness for real-world humanoid mobility.

Abstract

Humanoid perceptive locomotion has made significant progress and shows great promise, yet achieving robust multi-directional locomotion on complex terrains remains underexplored. To tackle this challenge, we propose RPL, a two-stage training framework that enables multi-directional locomotion on challenging terrains, and remains robust with payloads. RPL first trains terrain-specific expert policies with privileged height map observations to master decoupled locomotion and manipulation skills across different terrains, and then distills them into a transformer policy that leverages multiple depth cameras to cover a wide range of views. During distillation, we introduce two techniques to robustify multi-directional locomotion, depth feature scaling based on velocity commands and random side masking, which are critical for asymmetric depth observations and unseen widths of terrains. For scalable depth distillation, we develop an efficient multi-depth system that ray-casts against both dynamic robot meshes and static terrain meshes in massively parallel environments, achieving a 5-times speedup over the depth rendering pipelines in existing simulators while modeling realistic sensor latency, noise, and dropout. Extensive real-world experiments demonstrate robust multi-directional locomotion with payloads (2kg) across challenging terrains, including 20° slopes, staircases with different step lengths (22 cm, 25 cm, 30 cm), and 25 cm by 25 cm stepping stones separated by 60 cm gaps.

Figures (11)

• Figure 1: RPL enables long-horizon bidirectional locomotion and remain robust with payloads on Unitree G1 humanoid robot with a depth-based transformer policy. Terrains include 20° slopes, staircases with different step lengths (22 cm, 25 cm, 30 cm), and 25 cm$\times$25 cm stepping stones with 60 cm gaps.
• Figure 2: Overview of the Two-Stage Training Framework in RPL. Top (Stage 1): We train terrain-specific expert policies using privileged height-map observations to master decoupled locomotion and manipulation. Bottom (Stage 2): We distill the expert policies into a single multi-view, depth-based transformer policy. Multi-camera depth inputs are encoded by CNN backbones with random side masking (\ref{['sec:rsm']}) and depth feature scaling based on velocity commands (\ref{['sec:dfsv']}), fused into visual features, and combined with noisy proprioceptive observations and task goals to predict student actions.
• Figure 3: Stepping stones without (left) and with (right) foot edge penalty.
• Figure 4: Stairs without (left) and with (right) foothold penalty.
• Figure 5: Illustration of DFSV. CNN feature maps from four directional depth cameras under a zero command ($v_x{=}0, v_y{=}0$, left) and a diagonal command ($v_x{=}0.5, v_y{=}0.5$, right), where the heatmaps visualize each camera's relevance to the commanded velocity direction.