A microrobot navigation policy trained in simulation in under 10 minutes was transferred to two differently shaped magnetic microrobots without task-specific retraining, according to a study published in Nature Machine Intelligence on September 28. The physical tests used artificial vessel-like channels; the paper does not report tests inside an animal or person.
The policy receives image-based information about the robot, its target and nearby channel boundaries, then sends a movement command to magnetic-control hardware. A ray-based feature extractor, which the authors call “virtual LiDAR,” turns the camera image into a compact view of open space around the robot.
Training speed comes from simulating thousands of channels in parallel. The system used a dataset of 13,278 computer-generated vascular environments and produced about 190,000 simulated transitions per second. A transition is one observed state, the policy’s action and the simulator’s result. The authors report that this setup trained a navigation policy in under 10 minutes; that is simulation time, not a measure of how quickly a physical robot completes a route.
The training feedback also matters. Besides rewarding arrival at a target and penalizing collisions, the researchers added a progress signal that tells the policy whether each move brings it closer, plus regularization that favors smoother control and greater obstacle clearance. In the authors’ tests, progress feedback mainly sped convergence, while regularization improved trajectory-quality measures.
The same policy was tested on a roughly 20-micrometer, iron-oxide-coated wild-chrysanthemum pollen grain and a helical microrobot. Each design completed two recorded routes in one previously unseen artificial vascular channel. Both also avoided 11 moving obstacles that were absent from training. The authors describe this as zero-shot transfer because they did not retrain or adapt the policy for each task, but the small number of physical routes limits what the tests establish.
A cerebral-artery phantom test used a hybrid controller: the learned policy guided horizontal movement from a top view, while a separate proportional-derivative controller set the robot’s height using side-view feedback. This was a laboratory demonstration, not a single learned policy controlling all three dimensions or navigation in living tissue; the authors say realistic vascular conditions remain a challenge.
Sources
- Nature Machine Intelligence version of record
- Full paper, methods and Figure 5 (arXiv)
- Authors’ microrobot-navigation code
Image credit: Figure 5 from Sun et al., “Minute-scale training for microrobot navigation,” under CC BY 4.0 (via arXiv).