A quadruped robot just learned to parkour without ever seeing clean camera data.
Researchers built a system called DAWN (Denoising and Alignment in World models for Noise-robustness) that trains a robot's internal world model to expect noisy depth-camera input from the start, rather than relying on hand-tuned filters bolted on after training. The model receives noisy depth as input but is trained to reconstruct clean depth as its target, forcing it to denoise implicitly. A contrastive-learning step further aligns the noisy and clean latent representations so the robot's internal picture of the world stays consistent regardless of sensor noise. Tested zero-shot on a Unitree Go1, the approach let the robot climb stairs up to 18 cm, clear gaps up to 70 cm, and mount steps up to 45 cm using raw, unfiltered depth data.
Most vision-based legged robots depend on manually tuned filter parameters that researchers rarely publish, which makes results hard to reproduce and performance fragile once real-world sensor noise creeps in. DAWN bakes noise tolerance into training itself and adds no extra computational cost at inference, which matters for a field where lab demos routinely fall apart outdoors.
Whether this generalizes past one robot and a handful of obstacle types is still unproven, but it beats the field's usual fix of bolting on yet another filter.