É«É«À²

Title of the thesis: Learning versatile skills with sample-efficient reinforcement learning

Thesis defender: Yi Zhao
Opponent: Prof. Olov Andersson, KTH, Sweden
Custos: Prof. Joni Pajarinen, Aalto University School of Electrical Engineering 

This doctoral thesis investigates how robots can learn complex skills more efficiently using Reinforcement Learning (RL). Although RL enables robots to learn through experience, current methods often require millions of interactions, which makes real-world training slow, costly, and difficult.

The thesis develops three approaches to improve data efficiency. First, it introduces Temporal Consistency Reinforcement Learning (TCRL), which learns useful internal representations of a robot’s state and improves learning in locomotion tasks. Second, it develops methods for using previously collected data more effectively, including data of varying quality and data from different robotic systems. These methods improve training stability and sample efficiency across a wide range of visual motor-control tasks. Third, the thesis studies how better reward design can support the learning of highly complex skills. Using robotic piano playing as an example, a new reward formulation enables robots to learn coordinated two-handed performance. This research also resulted in large-scale robotic piano datasets containing more than one million trajectories and the development of OmniPianist, an agent capable of performing diverse musical pieces.

Overall, the thesis shows that better representations, effective use of existing data, and informative reward signals can significantly reduce the amount of new experience robots need to learn complex skills. The results provide practical tools for developing more efficient, stable, and versatile robotic learning systems.

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