Reinforcement learning in robotic granular materials

Collective behavior in biological systems—from the flocking of birds to the organizational structures in ant colonies—demonstrates how individuals can self-organize to exhibit complex emergent behavior. Prior approaches for designing physical systems with desired emergent properties rely on defining explicit interaction rules between the individual elements. However, as the number of interacting units increases, this approach becomes more difficult. In this work, we leverage recent advances in artificial intelligence, reinforcement learning, and swarm intelligence to design autonomous agents that can collaborate and self-aggregate into materials with predefined geometries and mechanical properties. The agents are designed to mimic grain-sized robots that, when collaborating, behave as pre-programmed robotic granular matter. We aim to develop efficient behavioral strategies that enable these agents to navigate, rearrange, and assemble in dense environments near the jamming limit. This work extends the frontier of active and intelligent matter beyond naturally occurring forms, offering new routes toward programmable materials.