DAWN: Noise-Robust Quadruped Parkour via Depth-Denoising World Models

Yohan Choi; Min-Jun Kim; Jin-Sung Kim; Yong-Jae Kim; Youn-Hee Han

cs.ROSep 24, 2026

DAWN: Noise-Robust Quadruped Parkour via Depth-Denoising World Models

Authors: Yohan Choi, Min-Jun Kim, Jin-Sung Kim, Yong-Jae Kim, Youn-Hee Han

Organizations: Convergence Engineering, Korea University of Technology and Education, Cheonan 31253, South Korea · School of Electrical, Electronics and Communication Engineering, Korea University of Technology and Education, Cheonan, 31253, South Korea

Abstract

Vision-based legged locomotion methods assume clean depth at training time and rely on hand-tuned post-processing filters at deployment. However, filter parameters are rarely disclosed, hindering reproducibility, and performance degrades substantially when depth noise is left unaddressed. Building noise robustness directly into the learning pipeline would eliminate this dependency. While such robustness has been explored for proprioceptive inputs, analogous approaches for depth perception remain largely absent in legged locomotion. We propose DAWN (Denoising and Alignment in World models for Noise-robustness), a noise-robust perception framework for legged locomotion, which builds noise robustness directly into a world model via two modifications: (1) feeding noisy depth to the encoder while keeping clean depth as the reconstruction target, forcing the model to implicitly denoise its input; and (2) applying contrastive learning to align the latent states of noisy and clean depth. Importantly, DAWN is not tied to a specific noise model, requiring no manual tuning to the noise distribution at deployment. Furthermore, it incurs no additional inference cost over existing world model-based methods. Without any manual filter calibration -- relying solely on the learned noise-robust representation -- DAWN achieves zero-shot quadruped parkour on a Unitree Go1: traversing stairs up to 18 cm, clearing gaps up to 70 cm, and mounting steps up to 45 cm from raw depth observations. Ablation studies show that denoising and contrastive alignment contribute at complementary levels -- reconstruction and representation, respectively -- and yield additive gains when combined. Videos and code are available at: https://dawn-parkour.github.io/

Figures & tables

Fig. 1: Overview of the DAWN framework. DAWN’s two modifications are highlighted in green: the denoising reconstruction loss $\mathcal{L}_{\text{denoise}}$ between the decoder output and the clean depth target, and the contrastive loss $\mathcal{L}_{\text{contrast}}$ applied to the projection head (MLP).

	WMP	DAWN
Encoder input	$x^{\text{c}}_{t}$	$x^{\text{n}}_{t}$
Decoder target	$x^{\text{c}}_{t}$	$x^{\text{c}}_{t}$
Reconstruction loss	$\mathcal{L}_{\text{WMP}}$	$\mathcal{L}_{\text{denoise}}$
Contrastive loss	–	$\mathcal{L}_{\text{contrast}}$
Post-processing filter	Required	Not required

TABLE I: Comparison of WMP and DAWN training configurations. The RSSM architecture and policy training are identical; only the encoder input and loss terms differ.

	Slope (6–22 ∘ )		Stair (6–18 cm)		Gap (10–90 cm)		Step (10–54 cm)
Method	SR	TE	SR	TE	SR	TE	SR	TE
Blind	$98.9{\pm 3.8}$	$0.025{\pm.004}$	$74.6{\pm 44.5}$	$0.064{\pm.066}$	$26.2{\pm 44.0}$	$0.154{\pm.073}$	$36.2{\pm 43.4}$	$0.936{\pm.071}$
EP	$100.0{\pm 0.0}$	$0.011{\pm.003}$	$82.2{\pm 35.4}$	$0.031{\pm.042}$	$83.5{\pm 35.9}$	$0.072{\pm.048}$	$92.6{\pm 23.2}$	$0.038{\pm.031}$
WMP	$100.0{\pm 0.0}$	$0.010{\pm.003}$	$89.7{\pm 30.4}$	$0.026{\pm.038}$	$88.6{\pm 31.8}$	$\mathbf{0.058}{\pm.034}$	$95.9{\pm 19.8}$	$0.026{\pm.018}$
WMP w/ N	$99.9{\pm 3.2}$	$0.009{\pm.003}$	$94.5{\pm 22.8}$	$0.023{\pm.036}$	$93.7{\pm 24.3}$	$0.067{\pm.029}$	$95.3{\pm 21.2}$	$\mathbf{0.025}{\pm.021}$
DAWN	$\mathbf{99.9}{\pm 2.5}$	$\mathbf{0.008}{\pm.002}$	$\mathbf{96.6}{\pm 18.2}$	$\mathbf{0.015}{\pm.020}$	$\mathbf{97.2}{\pm 16.6}$	$0.065{\pm.031}$	$\mathbf{97.0}{\pm 17.0}$	$0.034{\pm.020}$

TABLE II: Baseline comparison (noise

\times 1.0

). SR (%,

\uparrow

) and TE (m/s,

\downarrow

) are reported. Averaged over all difficulty levels, 3 seeds

\times

100 episodes.

Method	Noise	Denoise	Contrastive
Blind	–	–	–
EP [ 8 ]	×	×	×
WMP [ 9 ]	×	×	×
WMP w/ N	✓	×	×
DAWN w/o C	✓	✓	×
DAWN w/o D	✓	×	✓

TABLE III: Summary of compared methods. Noise: depth noise applied during training. Denoise: reconstruction target set to clean depth. Contrastive: contrastive learning applied.

Fig. 2: Simulation analysis. Ablation study showing average SR over Stairs/Gap/Step by difficulty level (noise ×1.0), mean of 3 seeds (left). Noise robustness analysis showing average SR over Stairs/Gap/Step at the highest difficulty as the noise scale increases from $\times 1.0$ to $\times 2.0$ , averaged over 3 seeds (center). t-SNE visualization of encoder latent states; colors indicate terrain type, shading indicates input condition (filled: clean, open: noisy) (right).

Fig. 3: Comparison of simulated clean depth, simulated noisy depth ( $\times 1.0$ ), and real D435i depth across three terrains (Stairs, Gap, Step). The simulated noise patterns closely match real sensor output.

Fig. 4: Real-world depth reconstruction on three terrains (Stairs, Gap, Step). For each terrain, the top row is the raw D435i depth input and the bottom row (dashed border) is DAWN’s decoder output at the same timestep, shown at t = k, 2k, …, 5k, 15k, and 25k.

Parameter	$\times 1.0$	$\times 1.5$	$\times 2.0$
Gaussian $\sigma$ (m)	0.01	0.015	0.02
Edge $P_{\max}$	0.60	0.80	0.90
Particle $r_{p}$	0.003	0.0045	0.006

TABLE IV: Noise scale parameters.

\times 1.0

is the default setting.

Fig. 5: Real-world experimental results. Three methods (DAWN, WMP w/ N, WMP) $\times$ three terrains (Stairs, Gap, Step) $\times$ two environments (Indoor, Outdoor). SR (%) over 10 trials per difficulty level.

Fig. 6: Outdoor deployment snapshots: curb climbing on stone terrain, stair descending, slope descending, gravel walking, grass walking, and slope ascending.

Explore similar work

Sep 21, 2026cs.RO

REDACT: Robust Perceptive Locomotion under Unseen Visual Corruption

Depth-conditioned locomotion policies have demonstrated impressive agile maneuvers, but can be steered to unpredictable actions when observations are outside their training distribution. Occlusion, invalid returns, sensor noise, and visual distractors can shift deployment observations away from nominal simulated depth. While synthetic sensor augmentation targets specified degradations, it does not by itself define behavior under corruption families omitted from training. To address gaps in training-time coverage, we present REDACT (Retaining Evidence Despite Artifacts for Continued Traversal), a teacher-student framework combining an improved visual encoder architecture, persistent feature masking, and a novel consensus-gating algorithm to retain useful depth information under unmodeled corruption. The gate uses approximate conformal calibration on clean observations alone, requiring no prior knowledge of the corruption type. Trained on clean simulated depth, REDACT retains useful visual information under unseen corruption, supporting higher traversal success than existing parkour baselines. Evaluation of depth augmentation across corruption families further shows that REDACT improves robustness where augmentation coverage is missing. Real-world trials demonstrate zero-shot transfer to structured and forested environments with unfamiliar scene content.

Natapat Kirdwichai, Tobias Driskell-Poole, Andrei Sontea +3

The authors are with the Faculty of Engineering and Physical Sciences and the Faculty of Environmental and Life Sciences, University of Southampton, United Kingdom.

Robotic Perception Agile Locomotion

Ask PDF

Sep 11, 2026cs.RO

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

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.

Hongjin Chen, Zijun Xu, Shihao Ma +8

Fudan University · TARS Robotics · Shanghai Innovation Institute +2

Proprioception Visuomotor Policy

Ask PDF

Jun 18, 2026cs.RO

SWAP: Symmetric Equivariant World-Model for Agile Robot Parkour

While latent world models enable the proactive predictions required for extreme parkour, their purely data-driven nature forces them to redundantly encode left-right symmetric interactions as independent patterns. This inflates the learning burden and hinders the capture of geometric regularities, restricting the latent space's efficiency for downstream policies. To address this, we propose SWAP, an end-to-end equivariant symmetric world model. This framework embeds symmetry directly into both the world model and the actor-critic networks. In real-world tests, the robot leaps across a 2.13 m gap and climbs a 1.63 m platform, breaking records for quadruped parkour. Furthermore, the framework exhibits robust geometric generalization to unseen mirrored terrains and exceptional zero-shot transferability across diverse outdoor environments. These results demonstrate that symmetry equivariance is an effective structural prior for pushing the physical boundaries of learned legged locomotion.

Kaixin Lan, Ze Wang, Hongyi Li +6

Center for X-Mechanics, Zhejiang University, Hangzhou, China · Mirrorme Technology Co., Ltd. Shanghai, China · ZJU-Hangzhou Global Scientific and Technology Innovation Center, Hangzhou, China

Scalable Robot Learning World Models Robot Systems Switch Soft Actor-Critic Data-Driven

Ask PDF