Researchers say a new training method helped a Unitree Go1 quadruped cross a 70-centimeter outdoor gap in six of 10 attempts by making its depth-camera system more tolerant of noisy images.
A depth camera estimates how far objects are from the robot, giving its controller a rough map of the ground and obstacles. Outdoors, sunlight can interfere with the camera’s infrared measurements, while different surface materials can add further noise. A controller trained on cleaner images may then misread a step or gap.
The method, called DAWN, changes training rather than adding a filter at deployment. It feeds noisy depth images into the robot’s perception model while asking that model to reconstruct a clean version. A second part of training teaches the model to treat noisy and clean images of the same scene similarly. The authors say DAWN adds no computation when the robot is running and avoids having to calibrate a different filter for each environment.
In the 70-centimeter outdoor gap test, DAWN succeeded in six of 10 trials. For comparison, the researchers’ policy trained with noisy input succeeded in four of 10 trials, while a version trained with clean input succeeded in three of 10. One additional success changes the reported rate by 10 percentage points.
The real-robot evaluation used one Go1 and an Intel RealSense D435i camera, with 10 trials at each difficulty level. The authors’ current sensor-noise model targets the D435i; whether the method transfers to other cameras or robot platforms remains untested.
Sources
- DAWN: Noise-Robust Quadruped Parkour via Depth-Denoising World Models, original paper.
- Official DAWN implementation.
- Illustrative image, not the study robot: photo by Krists Luhaers of a Unitree Go1 at Positivus, file record, licensed under CC BY 2.0.
.jpg){kind=link}