Research

We build agents that learn to perceive, act, and acquire skills in complex physical systems — uniting reinforcement learning, control and dynamical-systems theory, and information theory, with intrinsic motivation as a recurring thread.

Intrinsic Motivation & Empowerment

Information-theoretic objectives — empowerment, channel capacity, and the value of information — that drive agents to discover useful, generalizable behavior without hand-engineered rewards, including intrinsic motivation realized through deep reinforcement learning.

Information-Theoretic Perception & Control

The fundamental limits of perceiving and controlling dynamical systems: how much information an agent must acquire and retain in order to act, studied through information-bottleneck and channel-capacity analyses of closed-loop feedback.

Robot Learning & Skill Acquisition

Learning controllable skills on real physical systems — manipulation, locomotion, and sparse robotic actuation — with an emphasis on sample efficiency, stability (Lyapunov-guided learning), and transfer from simulation to hardware.

Learning in Contact-Rich Environments

Acting through frequent physical contact — dexterous manipulation, locomotion, and assembly — where intermittent contacts induce hybrid, non-smooth dynamics that challenge standard control and learning, and demand methods that reason about when and how to make contact.

Behavior Synthesis for Dynamical Systems

Synthesizing complex, structured behavior for nonlinear and hybrid dynamical systems by learning predictive, control-aware representations of dynamics that support planning and design.

Emergence & Multi-Agent Intelligence

How coordinated, complex collective behavior emerges in groups of agents from intrinsic objectives, rather than from centralized control or explicit reward design.

Foundations of Reinforcement Learning

The theory behind the applications above: entropy-regularized and average-reward RL, reward shaping and compositionality, and bounds on optimal value functions.