Training a Microrobot to Steer Itself Now Takes Minutes, Not Days

Researchers at the Hong Kong Polytechnic University report that they can train a microrobot to navigate on its own in under 10 minutes, a job that existing methods take hours to days to finish. The training happens entirely in simulation, inside a computer model holding more than 10,000 artificial blood vessels.

What the training produces is a policy: the rules the robot steers by, learned by trial and error rather than written out by a programmer. Writing in Nature Machine Intelligence, Yinghan Sun and colleagues say the speed-up would shorten the design loop for microrobots, machines small enough to be steered through narrow channels such as blood vessels.

The simulator runs those environments in parallel and reaches roughly 190,000 transitions per second. The team credits the fast learning to a reward scheme of its own design, which on their own measurements cuts variation in the robot's steering by at least 33.7% and widens the gap the robot keeps from obstacles by at least 2.1% in every scenario tested.

The same policies transferred to other types of microrobot and to new scenarios without further training. The paper also describes laboratory tests on real microrobots, scored on success rate, completion time and path efficiency. In one of them a robot navigated a benchtop model of human brain vasculature.

The vessel dataset used for training and the code behind the simulator are public, deposited on Zenodo and posted on GitHub.