Learning and Memory in Physical Matter

 

​Matter that learns: Our next question is whether these microscopic rules can be learned and generalized. We seek a shared local policy that runs independently on every particle and produces collective motion, reconfiguration, transport, and adaptation. Because every particle has the same sensing and action space, different robotic bodies can be composed by changing their number, arrangement, and connectivity. This creates a physical setting to test whether one policy can generalize across scale, morphology, environment, task, and changing physics (manuscript forthcoming).

​

Physical learning: Can matter learn through its own physical dynamics, without relying on externally trained neural networks? We investigate how particles can adapt their interactions using local physical signals, such as fluid flows and mechanical forces. By coupling learning rules directly to the physical medium, we aim to create materials that autonomously adjust their microscopic interactions to achieve desired collective behaviors.

​

The long-term vision is a "Large Matter Model": a foundation model that generates diverse physical structures and functions from a common vocabulary of intelligent particles. Just as large language models learn transferable rules over linguistic tokens, a Large Matter Model would learn transferable interaction rules over physical building blocks. By combining decentralized learning with programmable matter, we aim to create materials that can sense, adapt, and reorganize their own function. As the particles become smaller and more numerous, the distinction between a robotic swarm and an intelligent material begins to disappear.