Without retraining, these agents restore capillary flow to healthy baseline levels. That's the headline from a new arXiv preprint that trains deep RL agents to navigate a physically grounded simulation of blood capillaries, then proves they can pivot to intervention tasks they never saw during training.
The Simulation Is Not a Toy
The team built a capillary network simulation that includes realistic hydrodynamic flow fields, explicit red blood cell dynamics, and anatomically derived branching geometry. Prior RL studies of microrobotic navigation worked in idealized geometries that omitted complex flow, confined branching, and dense cellular obstacles. This work closes that gap. They train agents to navigate via chemotaxis, and systematically map the physical limits across robot size and swimming speed. One result: a forbidden regime where Brownian motion and flow overwhelm propulsion.
Discovery Without Engineering
Successful agents independently discover multiple universal strategy types. Run-and-rotate and energy-efficient search-and-sit policies emerge regardless of robot parameters. No hand-coded behaviors, no task-specific reward shaping. The agents find these strategies on their own.
Proof of transfer: without any retraining, these same agents perform targeted blocking and unblocking of capillary flow. They restore throughput to healthy baseline levels. That means a single training phase can generalize to intervention tasks that were not part of the original navigation objective.
Autonomous microrobots navigating biological vasculature could enable targeted drug delivery and thrombolysis. This work shows that RL is a viable framework for developing those intervention strategies in complex biological environments. The next step is moving from simulation to in vitro validation, but the path is now mapped.
Source: Reinforcement Learning Enables Autonomous Microrobot Navigation and Intervention in Simulated Blood Capillaries
Domain: arxiv.org
Comments load interactively on the live page.