Measuring the Robustness of Multi-Agent Reinforcement Learning Systems under Partial Agent Failure

Robust multi-robot behaviour hinges on more than network architecture: it critically depends on what the task punishes.The MAPPO algorithm has been analyzed on two canonical benchmarks that share a navigation objective but differ in their reward shaping.In Navigation, where collisions incur penalties, performance collapses once a critical fraction of agents is faulty, even under mild sensor noise.In Sampling, where collisions are ignored, the policy withstands swarm-wide degradation until the magnitude of observation noise exceeds the radius of the hotspot, after which the reward plummets uniformly for every agent ratio.The results indicate that the algorithm exhibits some robustness to certain types of perturbations, primarily due to the generalization properties of the neural networks, but this robustness is strongly shaped by the structure of the reward function.Three design lessons emerge: (i) reward shaping decides which disturbance axis is fatal; (ii) every task has a tolerance window where moderate degradation is harmless; (iii) failures are dominated by a handful of topologically central "hub" robots, so identifying and shielding these critical agents offers high leverage.Taken together, our findings chart a path towards resilient real-world deployments: combine collision-aware rewards with targeted protection of influence-maximising agents.