TAG - Testbed for Autonomous Games

TAG provides a compact physical platform for studying learning and control in physically coupled dynamical systems . The testbed consists of an autonomous maze board integrated with a rotary inverted pendulum¹. A dedicated controller actively tilts and steers the maze, while an overhead camera provides visual measurements of the marble's motion. At the same time, the top-mounted rotary inverted pendulum dynamically interacts with the shared platform, injecting reaction torque and vibration that couple the pendulum, board, and marble dynamics. This creates a challenging yet controllable environment in which multiple subsystems interact through shared physical dynamics rather than operating independently.

TAG supports research in model-based and model-free learning, system identification, reinforcement learning, adaptive and robust control, multi-agent decision-making, and learning under disturbances and partial observability. More broadly, our goal is to provide a reproducible physical benchmark for developing AI and control algorithms that must learn, predict, and make decisions in the presence of physical coupling, dynamic interactions, uncertainty, and nonstationary environments .

¹ The autonomous maze is inspired by CyberRunner.