Learning Robust Quadrupedal Locomotion via Asymmetric Adversarial Sabotage

The operational capability of legged robots in dynamic and unpredictable real-world environments depends heavily on their resilience and capacity to withstand severe external perturbations. To elevate the disturbance rejection and physical robustness of the Unitree Go2 quadruped robot to extreme levels, this study proposes an innovative asymmetric Adversarial Reinforcement Learning (ARL) framework founded upon the core principle of "learning through sabotage." Unlike conventional adversarial training paradigms that often rely on decoupled or overly complex network topologies, this framework uniquely features protagonist (controller) and antagonist (saboteur) agents that share a single, unified neural network architecture. This tightly integrated asymmetric structure enables the saboteur to rapidly discover and exploit the controller's latent vulnerabilities through direct parameter sharing, allowing it to apply precise, strategic perturbing forces precisely at the most critical junctures of the gait cycle. The overall training process is systematically executed via a meticulously designed three-stage pipeline: a warm-up phase dedicated to foundational locomotion acquisition, a development phase aimed at building baseline controller resilience, and a competitive zero-sum game phase where the intelligent saboteur relentlessly focuses on destabilizing the robot. Deliberately unconstrained by conventional trade-offs such as energy efficiency or mechanical power optimization, the optimization objective focuses exclusively on high-fidelity command tracking and dynamic stability under duress. Comprehensive stress tests conducted within the high-fidelity Isaac Sim environment demonstrate that the proposed asymmetric ARL method significantly increases the robot's physical tipping threshold by 50% compared to standard domain randomization techniques, while successfully preserving robust trajectory tracking and balance under extreme, unforeseen perturbations.