Online

Kriegman, Sam

Complex Systems and Data Science

2020

Ph. D.

Data-driven, rather than hypothesis-driven, approaches to robot design are becoming increasingly widespread, but they remain narrowly focused on tuning the parameters of control software (neural network synaptic weights) inside an overwhelmingly static and presupposed body. Meanwhile, an efflorescence of new actuators and metamaterials continue to broaden the ways in which machines are free to move and morph, but they have yet to be adopted by useful robots because the design and control of metamorphosing body plans is extremely non-intuitive. This thesis unites these converging yet previously segregated technologies by automating the design of robots with physically malleable hardware, which we will refer to as protean machines, named after Proteus of Greek mythology. This thesis begins by proposing an ontology of embodied agents, their physical features, and their potential ability to purposefully change each one in space and time. A series of experiments are then documented in which increasingly more of these features (structure, shape, and material properties) were allowed to vary across increasingly more timescales (evolution, development, and physiology), and collectively optimized to facilitate adaptive behavior in a simulated physical environment. The utility of increasingly protean machines is demonstrated by a concomitant increase in both the performance and robustness of the final, optimized system. This holds true even if its ability to change is temporarily removed by fabricating the system in reality, or by "canalization": the tendency for plasticity to be supplanted by good static traits (an inductive bias) for the current environment. Further, if physical flexibility is retained rather than canalized, it is shown how protean machines can, under certain conditions, achieve a form of hyper-robustness: the ability to self-edit their own anatomy to "undo" large deviations from the environments in which their control policy was originally optimized. Some of the designs that evolved in simulation were manufactured in reality using hundreds of highly deformable silicone building blocks, yielding shapeshifting robots. Others were built entirely out of biological tissues, derived from pluripotent Xenopus laevis stem cells, yielding computer-designed organisms (dubbed "xenobots"). Overall, the results shed unique light on questions about the evolution of development, simulation-to-reality transfer of physical artifacts, and the capacity for bioengineering new organisms with useful functions.