| Abstrakt: | This thesis studies how the NICO humanoid robot can learn useful behavior in a virtual Unity environment. The robot is trained with reinforcement learning through the ML-Agents toolkit. We first explore gaze control through a task where NICO learns to look at a selected object while ignoring distractors. Then we study pushing in a tabletop task where the robot moves cubes into target zones. The experiments show that strong policies require careful design of observations, rewards, exploration and curricula. The thesis also proposes a custom self-attention-based policy architecture for the pushing task, which learns the task successfully. The learned policies are then analyzed through ablations, counterfactual tests, local explanations, and attention weights in the self-attention-based policy.
|
|---|