This research proposes SWARM-LS, a framework for adaptive multi-robot coordination in partially observable environments. It combines a consensus-aware variational encoder, a graph-structured latent dynamics model, and an adaptive attention mechanism to address challenges in joint dynamics complexity, non-stationarity, and communication constraints.
Key findings
SWARM-LS enables decentralized multi-robot coordination through a shared latent space representation of environment dynamics.
The framework includes a consensus-aware variational encoder for mapping heterogeneous robot observations into a unified latent space.
A graph-structured latent dynamics model captures inter-robot dependencies through message passing.
An adaptive attention mechanism dynamically adjusts coordination intensity based on task requirements and communication bandwidth.
Limitations & open questions
The paper identifies potential failure modes including representation collapse, communication bottlenecks, and sim-to-real gaps.