Value-Aware MARO boosts multi-agent coordination by 20% under comms loss
New RL technique prevents performance collapse when communication drops below 40%
Multi-agent systems like drone swarms or autonomous vehicle platoons rely on constant communication, but real-world deployments frequently face intermittent failures. Traditional internal prediction models try to estimate missing data using standard reconstruction objectives, but they waste capacity on stochastic exploration noise and outdated suboptimal policies. Researchers from multiple institutions propose Value-Aware MARO (Multi-Agent Observation Sharing under Communication Dropout), which dynamically weights the predictor's loss function using advantage estimates derived from the underlying actor-critic architecture. This explicitly ties the predictor's learning to the policy's evolution, focusing model capacity on high-return dynamics.
Evaluated on several tasks in the Multi-Agent Particle Environment with varying communication reliability, the method maintains strong performance even below 40% reliability. It achieves an average improvement in mean returns of over 20% and reduces performance variance by a mean of 64.7% compared to the standard unweighted baseline. The paper has been accepted to the 2026 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), highlighting its significance for real-world robotic coordination under communication constraints.
- Value-Aware MARO uses advantage estimates to weight predictor loss, focusing on high-reward dynamics
- Achieves >20% improvement in mean returns and 64.7% reduction in performance variance over baselines
- Prevents performance collapse especially when communication reliability drops below 40%
Why It Matters
Enables robust drone swarms and autonomous vehicle fleets to maintain coordination during real-world communication failures.