Multi-Robot Traversal of 3D Terrain Using Adversarial Reinforcement Learning

Autonomous navigation of multi-robot systems in unstructured 3D terrain remains a critical challenge for search-and-rescue (SAR) operations. We present a novel adversarial reinforcement learning (ARL) framework that enables robust sim-to-real transfer for quadrupedal swarms traversing rubble-like environments. Our approach combines: ●       A multi-agent proximal policy optimization (MAPPO) baseline with adversarial perturbations ●       A domain randomization engine simulating debris variability (slope, friction, obstacle density) ●       Hardware-in-the-loop (HIL) validation using ANYmal-C robots Experimental results demonstrate a 92% success rate in unseen terrains, outperforming classical SLAM-based methods by 37%. This work bridges the sim-to-real gap for legged swarm robotics in disaster scenarios.