CAP: Continuously Adaptive Perception-Blind Humanoid Locomotion via Learned Denoising

Hongjin Chen, Zijun Xu, Shihao Ma, Yi Zhao, Xilai Liu, Ke Ma, Wei Zhang, Chunyang Xie, Pengfei Li, Jieru Zhao, Wenchao Ding · N/A · 2026

Summary

Humanoid locomotion across complex terrain demands forward-looking exteroception to anticipate obstacles, yet this signal is unreliable in real-world deployment, failing partially and intermittently.

Abstract Summary

Humanoid locomotion across complex terrain demands forward-looking exteroception to anticipate obstacles, yet this signal is unreliable in real-world deployment, failing partially and intermittently. Existing perceptive policies often assume that depth observations remain clean and in-distribution, while recent attempts to unify perceptive and blind control typically route or switch between separate sub-policies, leaving recoverable information in partially corrupted depth unexploited. We instead propose CAP, a single-stage humanoid locomotion policy that recovers this signal with a perceptive world-model encoder trained as a learned denoiser to reconstruct clean depth from a corrupted input, together with a co-active proprioceptive variational encoder that supplies depth-free body-state information. A coupled training recipe pairs a depth-noise curriculum on the world-model input with world-model feature dropout on the policy-facing latent, exposing the policy to failures across the entire perception-quality spectrum. In simulation, CAP matches or improves upon perceptive baselines when depth remains informative, and degrades more smoothly than a binary-switching baseline as perception worsens. On the Unitree G1, controlled trials and indoor-outdoor deployments demonstrate perception-robust locomotion under intermittent occlusion, real-sensor corruption, and outdoor depth artifacts.

Key Points

Humanoid locomotion across complex terrain demands forward-looking exteroception to anticipate obstacles, yet this signal is unreliable in real-world deployment, failing partially and intermittently.
Existing perceptive policies often assume that depth observations remain clean and in-distribution, while recent attempts to unify perceptive and blind control typically route or switch between separate sub-policies, leaving recoverable information in partially corrupted depth unexploited.
We instead propose CAP, a single-stage humanoid locomotion policy that recovers this signal with a perceptive world-model encoder trained as a learned denoiser to reconstruct clean depth from a corrupted input, together with a co-active proprioceptive variational encoder that supplies depth-free body-state information.
A coupled training recipe pairs a depth-noise curriculum on the world-model input with world-model feature dropout on the policy-facing latent, exposing the policy to failures across the entire perception-quality spectrum.
In simulation, CAP matches or improves upon perceptive baselines when depth remains informative, and degrades more smoothly than a binary-switching baseline as perception worsens.

Method

Fig. summarizes CAP ’s architecture and training recipe. Architecturally, CAP couples an asymmetric actor–critic with two co-trained encoders: a perceptive world-model encoder and a proprioceptive variational encoder. The asymmetric design [ 23 ] lets us collapse training into a single stage: the actor receives only signals available on the deployed robot, whereas the critic additionally accesses privileged signals.

Experiments

All simulation experiments use the Unitree G1 humanoid in Isaac Gym [ 21 ] with a common reward design on Stair, Platform, Gap, and Hurdle. We evaluate two anchor cells, Clean and Noisy , and an eight-stage perception-quality sweep from nominal depth to severe failure, with the severe endpoint tested only on Stair.

Each Noisy or degraded stage fixes the severity level, but its depth corruptions remain time-structured within a rollout through per-frame noise, occlusion, and intermittent full-frame failure.

Sources and Demonstrations

Figures are reproduced from Chen et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.