Terrain Adaption Controller for a Walking Excavator Robot using Deep Reinforcement Learning

Abstract

Automation of heavy-duty vehicles using technologies developed in the robotics domain is gaining popularity. One such vehicle is the walking excavator with active suspension chassis for adapting to uneven terrain. The terrain adaption controller automates the suspension control by considering the factors stability, underlying terrain structure, wheel-ground distance, chassis-ground distance, etc. This work builds the controller, which actuates the joints that control the height of the wheels. Deep reinforcement learning is used, considering the complexity of the problem and transferability to other robots. The controller is learned and evaluated in simulation, where continuous terrain with varying slopes is automatically generated. Autoencoders compress the height-map and convert it into latent space of different sizes. Three groups of controllers are then designed based on observations given to the controller. Evaluation of controllers shows that the controllers with ground distances as observation perform better. If the ground distances are part of the observation, there is no significant difference in performance between controllers with different latent space sizes. For controllers with terrain information and no ground distances, the evaluation results match the terrain reconstruction accuracy of the corresponding autoencoder.