Learning Robust Perception-Based Controller for Quadruped Robot

The perception of a quadruped robot is crucial in determining how the robot has to move while crossing challenging terrains. However, robustness in the quadruped perception is still a dilemma when facing a slippery, deformable, and reflective terrains. This problem arises due to the presence of noises from slippage on the foothold and misinterpretation of the terrain. The noises can elicit imperfect robot states and lead to the poor performance of the robot locomotion. This paper presents a robust perception-based control policy as a solution to overcome noises in the robot's perception. In particular, attention is paid to devise a robust perception method against noises both in proprioceptive and exteroceptive observations. In other words, the proposed control policy has a capability to estimate states and reduce effect of noises from both observations in the robot. A student-teacher algorithm is leveraged to train the control policy on a randomly generated terrain. Multiple robots also are employed in parallel learning regimes to reduce the training duration. The result is a robust control policy that can produce the optimal actions even when the robot's perception is affected by the noises observed not only the proprioceptive but also exteroceptive sensors. The robust control policy can handle a higher ratio of noise compared to the previous works in the literature. Robustness of the proposed control policy is validated using both the Isaac Gym simulation and a comparison with popular recurrent networks in the literature.

Automated summary

This paper presents a robust perception-based control policy to overcome noises in the robot's perception when crossing challenging terrains. The control policy can estimate states and reduce the effect of noises from both proprioceptive and exteroceptive observations, making it capable of handling a higher ratio of noise. The robustness of the control policy is validated through simulations and comparisons with recurrent networks.