Robot Thermodynamics.

다국어 초록 (Multilingual Abstract)

The pursuit of precision has been a driving force in engineering since the earliest days of the steam engine. Robotics, born from industrial automation, has embraced this focus. Robots are designed for low tolerances, absolute repeatability, and predictable behavior. Any uncertainty-in the environment, perception, or movement-is seen as a problem to be eliminated. This approach stands in contrast to the messy, unpredictable inner workings of biological organisms. Yet, despite these "flaws," living beings possess a degree of autonomy no machine can match. While precise determinism has its place in engineering, its blind pursuit limits the development of true "life-like" autonomy. This thesis explores a framework that embraces noise and uncertainty as essential tools, rather than obstacles, on the path toward more adaptable, and reliable, autonomous systems.This thesis proposes design, learning, and control principles for embodied agents with robust, nondeterministic, autonomy. It draws inspiration from (and contributes to the literature of) statistical mechanics and thermodynamics to produce results applicable to nonequilibrium systems such as robots and living organisms. Thermodynamics describes the flow of energy through matter, and how this flow and its fluctuations can be harnessed to produce work. Analogously, this thesis-titled Robot Thermodynamics-investigates how actions are materialized by robot bodies, and how the fluctuations induced by these actions can affect an agent's task-capabilities. In this endeavor, our primary unit of analysis is the path or trajectory distribution, which describes all possible paths through time and space that an agent can traverse. The structure of an agent's path distribution depends on its physical or material properties, as well as its controller or policy. Exploiting the relationship between agent behavior, embodiment, and decision-making through design, learning, and control is the explicit goal of robot thermodynamics.This thesis begins by laying the analytical foundations of robot thermodynamics. This mathematical overview serves multiple purposes: First, it introduces the principle of maximum caliber as an inference framework over path distributions. Then, it illustrates how these inferred path distributions and their properties can be used to characterize and manipulate the dynamics of complex systems. Lastly, it describes how optimal control and reinforcement learning can be framed as operations applied onto an agent's path distribution. The thesis then proceeds by demonstrating the power of this approach in several different applications across length-scales-prediction and control of nonequilibrium collectives, design of energy-harvesting colloidal microparticles, and embodied reinforcement learning-each advancing the state-of-the-art in their respective fields. Taken together, the results in this thesis highlight the promise of noise and uncertainty as versatile tools in the development of robust, life-like, real-world autonomy.