Locomotion Beyond Feet

TL;DR

The paper tackles the problem of enabling stable, whole-body humanoid locomotion on obstacle-rich environments where leg-only strategies fail. It presents a four-component framework that fuses physics-grounded keyframe motions with DeepMimic-style motion tracking, a depth-conditioned visual skill classifier, and a hierarchical policy execution that unifies vision and proprioception. Key contributions include a seven-component system for diverse terrains, a real-time depth-based skill classifier trained on real-world data, and extensive zero-shot, sim-to-real demonstrations across low-clearance spaces, walls, platforms, and stairs, validated on a compact humanoid. The work demonstrates robust, generalizable, contact-rich locomotion with practical implications for real-world deployment and transfer to full-size humanoids, while outlining avenues for automated keyframe design and more accurate contact modeling.

Abstract

Most locomotion methods for humanoid robots focus on leg-based gaits, yet natural bipeds frequently rely on hands, knees, and elbows to establish additional contacts for stability and support in complex environments. This paper introduces Locomotion Beyond Feet, a comprehensive system for whole-body humanoid locomotion across extremely challenging terrains, including low-clearance spaces under chairs, knee-high walls, knee-high platforms, and steep ascending and descending stairs. Our approach addresses two key challenges: contact-rich motion planning and generalization across diverse terrains. To this end, we combine physics-grounded keyframe animation with reinforcement learning. Keyframes encode human knowledge of motor skills, are embodiment-specific, and can be readily validated in simulation or on hardware, while reinforcement learning transforms these references into robust, physically accurate motions. We further employ a hierarchical framework consisting of terrain-specific motion-tracking policies, failure recovery mechanisms, and a vision-based skill planner. Real-world experiments demonstrate that Locomotion Beyond Feet achieves robust whole-body locomotion and generalizes across obstacle sizes, obstacle instances, and terrain sequences.

Figures (7)

• Figure 1: Locomotion Beyond Feet enables whole-body humanoid locomotion on diverse and challenging terrains—including low-clearance spaces under chairs, knee-high walls, knee-high platforms, and steep ascending and descending stairs—through chaining nine distinct locomotion skills that actively engage body parts beyond the legs for stability and support.
• Figure 2: System Pipeline. First, we generate physics-grounded keyframe motions with a physics-aware GUI application, where robot poses and arrival times are specified interactively. Second, we interpolate the keyframes to create reference motions, which serve as tracking rewards for RL policies. We further apply extensive domain randomization, such as initial robot states, obstacle dimensions, frictions, and IMU noise. Finally, a skill planner processes depth input from a learned depth estimation module at $3.1~\mathrm{Hz}$, along with IMU readings and the current skill, to select the next appropriate skill.
• Figure 3: Test Obstacles. We show the robot beside test obstacles, including (a) low-clearance spaces under chairs, (b) knee-high platforms, (c) knee-high walls, and (d) steep ascending and descending stairs. The space under the chairs is shorter than the robot ($53~\mathrm{cm}$), requiring crawling. The wall is 48% of the robot's leg length ($25~\mathrm{cm}$), requiring climbing. The platform height is 44% of the leg length, and each stair height is 16% of the leg length, all posing extreme challenges at the robot’s scale.
• Figure 4: Motion Tracking Policies. We demonstrate our policies on traversing extremely challenging terrains—including (a) low-clearance spaces under chairs, (b) knee-high walls, (c) knee-high platforms, and (d) steep ascending and descending stairs—and additionally show (e) fall recovery from supine and prone positions in case of failure.
• Figure 5: Sim-to-real Depth Comparison. We set up the same scene of YCB objects calli2015ycb (a) in MuJoCo todorov2012mujoco and (b) in the real world. The real-world RGB images are rectified after calibrating the fisheye cameras’ intrinsics and distortion, with white dashed lines illustrating proper alignment. (c) On the right is a comparison of ground-truth depth with real-world estimates from Foundation Stereo wen2025foundationstereo with resolution $480\times640$ and $192\times256$, respectively. We compute the quantitative results in the cropped region marked by the black box.