Learning to control self-assembling morphologies: a study of generalization via modularity

Deepak Pathak, Chris Lu, Trevor Darrell, Philip Isola, Alexei A. Efros

Abstract: Much of contemporary sensorimotor learning assumes that one is already given a complex agent (e.g., a robotic arm) and the goal is to learn to control it. In contrast, this paper investigates a modular co-evolution strategy: a collection of primitive agents learns to self-assemble into increasingly complex collectives in order to solve control tasks. Each primitive agent consists of a limb and a neural controller. Limbs may choose to link up to form collectives, with linking being treated as a dynamic action. When two limbs link, a joint is added between them, actuated by the 'parent' limb's controller. This forms a new 'single' agent, which may further link with other agents. In this way, complex morphologies can emerge, controlled by a policy whose architecture is in explicit correspondence with the morphology. In experiments, we demonstrate that agents with these modular and dynamic topologies generalize better to test-time environments compared to static and monolithic baselines. Project videos are available at https://doubleblindICLR19.github.io/self-assembly/

Keywords: modularity, compostionality, graphs, dynamics, network

TL;DR: Learning to control self-assembling agents via dynamic graph networks

Community Implementations: [ 3 code implementations](https://www.catalyzex.com/paper/learning-to-control-self-assembling/code)