MASTARS: Multi-Agent Sequential Trajectory Augmentation with Return-Conditioned Subgoals
Summary
- Researchers introduce MASTARS, a diffusion-based framework designed to tackle the diversity-coordination trade-off in offline multi-agent trajectory augmentation.
- MASTARS generates multi-agent trajectories sequentially across agents while maintaining cross-agent consistent coordination using inpainting.
- Experiments across benchmarks like MPE, SMAC, and SMAC-v2 show that MASTARS delivers significant performance gains across multiple offline MARL methods.
Why it matters
This research addresses a fundamental bottleneck in offline multi-agent reinforcement learning by enabling the generation of diverse yet globally coherent trajectories, improving overall model performance on standard benchmarks.