Browse Teach Curate

Extrinsic Symmetries and the Group Equivariant Graphormer for Multi-Agent RL

A framework that imposes extrinsic (externally constructed) group symmetries on multi-agent system dynamics so that equivariant, symmetry-enhanced learning applies even when a system has little intrinsic symmetry. Its core architecture is the Group Equivariant Graphormer, a group-modular graph-transformer policy for distributed swarm control, validated on symmetry-breaking quadrotor swarms with improved generalization and zero-shot scalability to varying swarm sizes.

Wall-E Text

Proceedings of Machine Learning Research vol 283:1–26, 2025 Symmetries-enhanced Multi-Agent

Multi-agent reinforcement learning struggles with generalization, scalability, and sample efficiency, and a promising remedy is to bake a system's intrinsic symmetries into the policy so that equivar…