Skip to content
WE BUILDHUMANOIDS

RESEARCH

CAP keeps humanoid locomotion working when depth sensing fails (CoRL 2026)

Unitree G1LocomotionSim-to-realComputer Vision
ORIGINAL SOURCE · arXiv preprintRead on arXiv arxiv.org/abs/2609.11553

SUMMARY

CAP is a single locomotion policy for humanoids that keeps crossing complex terrain when its depth input is partly or intermittently corrupted. A learned denoiser reconstructs clean depth from the corrupted signal, a proprioceptive encoder supplies body-state information that does not depend on depth, and training exposes the policy to the full range of perception quality. The authors report that it degrades more smoothly than a switching baseline in simulation and demonstrate it on a Unitree G1 indoors and outdoors. The paper is accepted at the Conference on Robot Learning 2026.

Drafted with AI assistance from the source and reviewed by WBH. Follow the source link for the full text.

WHY IT MATTERS

EDITORIAL

Depth sensing fails in ordinary deployments through occlusion, sensor corruption and outdoor artefacts; one policy that degrades gracefully avoids hand-built switching between perceptive and blind controllers.

RELATED ON WBH