FastFunnels guides groups of racing robot cars through narrow gaps at speed

Vignesh Rajagopal and Rohan Chandra have described FastFunnels in an arXiv paper, a method for several car-like robots to move fast through cluttered, narrow spaces. Instead of fixed formations that change shape by set rules, it takes its cue from a water droplet: a reinforcement learning policy shapes a flexible elliptical patch that stretches, squeezes and turns around obstacles in real time.

Each robot is then steered by distributed model predictive control to stay inside the patch while avoiding the others. Tested on simulated F1TENTH racing cars in 14 scenarios with one to six robots, FastFunnels completed 89% of runs at 7.73 m/s, while the other methods either failed to finish or managed only very low speeds of 0.07 to 1.79 m/s, the authors report.