Physical Atari: A Robust and Accessible Platform for Real-time Reinforcement Learning on Robots

The article introduces "Physical Atari," a system designed for real-time reinforcement learning on robots, comprising the Robotroller, which actuates an Atari CX40+ controller, and an Atari Devbox for rendering game frames and reward signals. Built with affordable, off-the-shelf components and 3D-printed parts, the system can be assembled for under $1,000 and has demonstrated durability through extensive testing, highlighting the significance of on-device adaptation to mitigate performance degradation due to distribution shifts. This platform provides researchers with a robust and accessible means to study reinforcement learning algorithms in practical robotic applications.