Harnessing Physical Learning Theory to Multiplex Actuation and Sensing in Soft Robots

There is an emerging desire to embed decision making and learning directly into materials. In soft robotics, it is believed this can improve autonomy without needing electronic control systems or by reducing the computational power needed in a central processor. The most common approaches for achieving this physical learning implement physical logic functions in the robot body or mimic a machine learning approach for achieving optimal weights of connections in a network. In the later, the neurons or weights adjust according to global calculations of a central processor, but in the brain neurons and synapses adjust themselves using only local information. Recent work has demonstrated man-made self-adjusting and distributed systems capable of performing machine-learning problems,¹ which has promising scaling advantages over typical neural networks in power consumption, speed, and robustness to damage. These have been demonstrated in transistor-based self-adjusting analog networks that trains themselves to perform a wide variety of tasks, using a process call coupled learning.<br/><br/>This talk introduces our work on applying coupled learning in fluidic networks instead of the electronic networks of prior work. Our goal is to combine the robust manufacturing approaches for integrating fluid networks in soft robots with coupled learning theory to allow the robots to learn to adjust their physical properties towards a desired goal in response to environmental stimuli. We will introduce the theoretical challenges of applying couple learning networks to fluid networks due to their non-linear relationship between flux and pressure. We will also introduce our fluidic network testbed and model, demonstrate how coupled learning can optimize a flow network to achieve a desired set of steady pressures in the network nodes, which we use to multiplex actuation and design new sensing strategies. This work could enable soft robots to circumvent the limitations of traditional material and fluidic network design choices.<br/><br/>Dillavou, Sam, et al. "Circuits that train themselves: decentralized, physics-driven learning." <i>AI and Optical Data Sciences IV</i>. Vol. 12438. SPIE, 2023.