Microrobots Could Learn to Find Their Way in Minutes Instead of Days

The slow part of building a robot too small to see is not the building. It is the waiting. The job these machines are meant to do is hard enough to picture: steer something invisible through a maze of branching tubes, in real time, with magnetic fields. Teaching one to find its own way has meant deep reinforcement learning, the standard method, and that has typically taken hours to days for a single navigation policy.

A team from The Hong Kong Polytechnic University, The Chinese University of Hong Kong and Harbin Institute of Technology has cut that wait to under 10 minutes. The work, led by Lidong Yang with first author Yinghan Sun, was published on September 28, 2026 in Nature Machine Intelligence. It is about three weeks old, so treat it as a recent result rather than this week’s news.

Ten thousand practice courses at once

Reinforcement learning works by letting a program try things, rewarding good outcomes and repeating, over and over. The bottleneck is how fast those tries can be served up. The team built a simulator that runs more than 10,000 artificial vascular environments, branching mazes that stand in for vessels, and handles the physics, the robot’s simulated view and feasibility checks for thousands of them at once. That comes to roughly 190,000 practice steps every second. According to the team’s preprint, training ran on a desktop computer with a single consumer graphics card.

Speed alone can produce a sloppy navigator, so they also redesigned the rewards the program learns from. They call it a task-shaping-regularization reward. According to the paper, it helped training settle faster, reduced jerky changes in the robot’s actions by at least 33.7 percent and increased how far it kept from obstacles by at least 2.1 percent across every scenario they tested. The first gain is large; the second is small.

From the simulator to real microrobots

The authors report that policies trained this way could be deployed, without retraining, across different types of microrobots and navigation scenarios. The preprint shows this in the lab with two real, magnetically steered microrobots: a particle built from a pollen grain and a tiny helical robot. They found their way through vessel-like channels, past moving obstacles and against flowing liquid, none of which they had seen in training. The team also guided the helical robot through a model of a brain artery, using the learned policy as part of a combined controller.

Why minutes instead of days matters

When training takes days, every change to a robot’s design means another long wait to find out how it steers. The authors write that minute-scale training can substantially shorten that design loop, letting engineers try idea after idea instead of waiting on each one.

The practice courses are shaped like blood vessels for a reason. The paper names targeted drug delivery and therapy inside blood vessels as uses of particular interest for microrobots. But nothing here has been tested in a living body, and the authors acknowledge that real blood vessels remain challenging because conditions inside them are so changeable and uncertain. That future is years away, and it depends on progress in materials, imaging and safety, not just software.

What this work changes right now is the pace of the people building these machines. An engineer can find out whether a new design can steer in about the time of a coffee break, rather than waiting hours or days to learn that it cannot. Faster cycles of trial and error are often how hard engineering problems give way, one small design at a time.