At the University of British Columbia, three engineering physics students have engineered an air‑hockey robot that never required a single session of physical practice. Instead, the entire learning process unfolded inside a high‑fidelity computer simulation, where every nuance of table tilt, puck bounce and surface irregularity was measured and reproduced.
The simulation was meticulously calibrated to keep positional error down to nearly one millimetre across the entire playing surface. A single camera, fitted with retroreflective tape on the puck and a distinctive marker on the mallet, captured motion at 120 frames per second, while a contour tracker kept the puck visible even when the gantry briefly obscured it.
Precision Engineering Behind the Scenes
The hardware relies on two belt‑driven motors controlled by an STM32 Blue Pill microcontroller, with a supercapacitor smoothing voltage spikes during rapid accelerations. Custom firmware translates voltage signals into third‑order transfer functions, enabling feedforward control complemented by PID feedback to keep the mallet on target.
On the software side, the team vectorized their analytical model of puck and mallet motion, allowing a typical Intel i5 laptop to run thousands of game instances simultaneously — approximately 230 times faster than real time. A small neural network with 112 parameters predicts the departing velocity, angle and uncertainty of the puck, while the Soft Actor‑Critic reinforcement‑learning algorithm trains policies with roughly 200,000 parameters.
When the learned policy was transferred to the real robot, it faced human opponents without any additional fine‑tuning. The system maintained synchronized round‑trip delays within a 60‑hertz loop, though minor deviations appeared compared to the simulated environment.
The project, spanning roughly two years and led by Hudson Nock, Ian Hartley and Mauro Ferraz, demonstrates how deep simulation can bridge the gap to physical robotics, delivering a compelling glimpse of autonomous, millimetre‑precise sport‑playing machines.