Scaling Swarm Coordination with GNNs—How Far Can We Go?

摘要

The scalability of coordination policies is a critical challenge in swarm robotics, where agent numbers may vary substantially between deployment scenarios. Reinforcement learning (RL) offers a promising avenue for learning decentralized policies from local interactions, yet a fundamental question remains: can policies trained on one swarm size transfer to different population scales without retraining? This zero-shot transfer problem is particularly challenging because the traditional RL approaches learn fixed-dimensional representations tied to specific agent counts, making them brittle to population changes at deployment time. While existing work addresses scalability through population-aware training (e.g., mean-field methods) or multi-size curricula (e.g., population transfer learning), these approaches either impose restrictive assumptions or require explicit exposure to varied team sizes during training. Graph Neural Networks (GNNs) offer a fundamentally different path. Their permutation invariance and ability to process variable-sized graphs suggest potential for zero-shot generalization across swarm sizes, where policies trained on a single population scale could deploy directly to larger or smaller teams. However, this capability remains largely unexplored in the context of swarm coordination. For this reason, we empirically investigate this question by combining GNNs with deep Q-learning in cooperative swarms. We focused on well-established 2D navigation tasks that are commonly used in the swarm robotics literature to study coordination and scalability, providing a controlled yet meaningful setting for our analysis. To address this, we introduce Deep Graph Q-Learning (DGQL), which embeds agent-neighbor graphs into Q-learning and trains on fixed-size swarms. Across two benchmarks (goal reaching and obstacle avoidance), we deploy up to three times larger teams. The DGQL preserves a functional coordination without retraining, but efficiency degrades with size. The ultimate goal distance grows monotonically (15–29 agents) and worsens beyond roughly twice the training size (≈20 agents), with task-dependent trade-offs. Our results quantify scalability limits of GNN-enhanced DQL and suggest architectural and training strategies to better sustain performance across scales.