Physical Learning in Soft Fluidic Channels through Experience

Fluidic circuits have received increasing interest as a paradigm for implementing computational functionality, e.g., for control in soft robots. However, typically control policies are encoded in circuits that are static, and only a few attempts have been made to realise physical learning capabilities that enable animal-like, real-time adaptation and lifetime development. We introduce the Fluidic Learning Channel (FLC), a physical learning framework that changes flow-conductance through its experienced flow rate history, allowing fluidic circuits to conduct physically embodied online learning. Two demonstrations are presented to validate this concept. The first involves a two-finger system that learns to memorise actuation speed under repetitively applied physical constraint. The second demonstrates a 2$\times$2 FLC network that learns to map the flow rate at the two input nodes to the target pressure at one of the output nodes. In addition to physical demonstration, simulations were conducted to further explore the essential characteristics and provide insights for future FLC-type embodiment designs.