Learning from Demonstration with Probabilistic Movement Primitives on Komatsu Excavator

Abstract

Automating heavy machine operation is a complex and challenging industry problem which demands constant improvement. As a case study on the topic, this thesis studies how learning can be utilized in industrial automation via a learning from demonstration implementation on the Komatsu PC138US-8 hydraulic excavator, to realize automation on free space movements. The implementation uses Probabilistic Movement Primitives (ProMPs) to determine a probability distribution over joint trajectories, whose samples are then executed with a trajectory follower implementation. This approach allows learning algorithms to learn the ProMP parameters, who capture both the demonstrated behavior and the associated variability which implies existence of properties for adjusting and generalization which makes the method an optimal choice for a case study.

The study considers a limited set of demonstrations, performing a digging operation cycle in free space movement, recorded in real life. An Expectation-Maximization algorithm is utilized to learn the ProMPs parameters from the dataset and the trajectory follower is implemented in MATLAB Simulink, which derives the feedforward calculation from the actuator dynamics and calculates the feedback signal with respect to the sampled trajectory.

Results show that ProMPs framework is able to operate the excavator by capturing different dynamics within the system on test environments. However, extending to more realistic, non-ideal cases require further work to solve limitations present within the theoretical framework, such as time and basis function dependencies. Some ideas for improvements are also discussed.