Public defence in Robotics and Autonomous Systems, M.Sc. Wenshuai Zhao

The title of the thesis: Robot learning across agents: from imitation to multi-agent cooperation

Thesis defender: Wenshuai Zhao
Opponent: Prof. Ville Hautamäki, University of Eastern Finland
Custos: Prof. Joni Pajarinen, Aalto University School of Electrical Engineering

As robots enter daily environments, they must learn both individual skills and coordinated teamwork. This thesis develops new learning methods that help robots acquire complex behaviors on their own and collaborate effectively in groups.

The thesis first addresses a key challenge in teaching humanoid robots from human motion data: human movements often violate the robot’s physical limits. The thesis introduces a new imitation learning framework that adjusts the robot’s target motion and control policy jointly, enabling robots to learn challenging skills such as jumping and kicking in a physically consistent way.

The research then focuses on multi-agent cooperation. Several new algorithms are proposed to overcome common failures in multi-agent reinforcement learning: avoiding suboptimal joint decisions, improving the efficiency of learning, and mitigating partial observation problems. A new curriculum strategy is also introduced to help robots learn more effectively in sparse-reward tasks.

Finally, the thesis connects single-agent and multi-agent learning by analysing how limited observations affect performance, showing that partial information can sometimes make learning more robust.

These results offer both new algorithms and insights for building autonomous systems that can learn complex skills and work together reliably, which may interest readers working in robotics and in fields such as autonomous driving and multi-agent systems.

Key words: reinforcement learning, imitation learning, multi-agent learning

Thesis available for public display 7 days prior to the defence at Aaltodoc. 

Contact: wenshuai.zhao@aalto.fi