Researchers led by Bao Trung Truong have built TAG, the Testbed for Autonomous Games, a small and cheap robotic platform for studying learning and control in systems that physically disturb each other, according to a paper on arXiv. It combines a tilting, steerable ball-in-maze board with a rotary inverted pendulum mounted on top.

The pendulum's reaction torque and vibration shake the board and the marble, while the board's tilting, driven by the policy that solves the maze, upsets the pendulum's balance. The team learned the maze controller with a model-based reinforcement learning method and the pendulum controller with a model-free one.

When the pendulum was switched on, the maze policy's success rate fell at once, but it adjusted while running and regained its previous level after a few episodes. In experiments, both policies, trained separately, controlled the board and the pendulum despite the coupling.

Source: arXiv Robotics (cs.RO)

Background

Bao Trung Truong

Robotics researcher

Robotics researcher.