Chemotactic navigation in robotic swimmers via reset-free hierarchical reinforcement learning

Microorganisms have evolved diverse strategies to propel themselves in viscous fluids, navigate complex environments, and exhibit taxis in response to stimuli. This has inspired the development of miniature robots, where artificial intelligence (AI) is playing an increasingly important role. Can AI endow these synthetic systems with intelligence akin to that honed through natural evolution? Here, we demonstrate, in silico, chemotactic navigation in a multi-link robotic model using two-level hierarchical reinforcement learning (RL). The lower-level RL allows the model-configured as a chain or ring topology-to acquire topology-adapted swimming gaits: wave propagation characteristic of flagella or body oscillation akin to an amoebae. Such chain and ring swimmers, further enabled by the higher-level RL, accomplish chemotactic navigation in prototypical biologically relevant scenarios that feature conflicting chemoattractants, pursuing a swimming bacterial mimic, steering in vortical flows, and squeezing through tight constrictions. Additionally, we achieve reset-free RL under partial observability, where simulated robots rely solely on local scalar observations rather than global or vectorial data. This advancement illuminates potential solutions for overcoming persistent challenges of manual resets and partial observability in real-world microrobotic RL.