Adaptation of a wheel loader automatic bucket filling neural network using reinforcement learning

Bucket-filling is a repetitive task in earth-moving operations with wheel-loaders, which needs to be automated to enable efficient remote control and autonomous operation. Ideally, an automated bucket-filling solution should work for different machine-pile environments, with a minimum of manual retraining. It has been shown that for a given machine-pile environment, a time-delay neural network can efficiently fill the bucket after imitation-based learning from 100 examples by one expert operator. Can such a bucket-filling network be automatically adapted to different machine-pile environments without further imitation learning by optimization of a utility or reward function? This paper investigates the use of a deterministic actor-critic reinforcement learning algorithm for automatic adaptation of a neural network in a new pile environment. The algorithm is used to automatically adapt a bucket-filling network for medium coarse gravel to a cobble-gravel pile environment. The experiments presented are performed with a Volvo L180H wheel-loader in a real-world setting. We found that the bucket-weights in the novel pile environment can improve by five to ten percent within one hour of reinforcement learning with less than 40 bucket-filling trials. This result was obtained after investigating two different reward functions motivated by domain knowledge.