Title: Optimizing the Design of Robot Environments via Interleaved Optimization and White-Box Motion-Planning
Advisor: Siddhartha Srinivasa
Supervisory Committee: Siddhartha Srinivasa (Chair), Blake Hannaford (GSR, Electrical and Computer Engineering), Adriana Schulz, Dieter Fox, Oren Salzman (Technion - Israel Institute of Technology)
Abstract:
Motion planning algorithms lie at the heart of all robotic systems. While each planner optimizes a different utility function, the robot’s environment is consistently the key factor affecting said utilities. Furthermore, in many scenarios (such as an assembly line) this environment is under human control. This leads to our first key insight: rather than treating a motion planner simply as a tool to be applied in a challenging environment, we can instead use the utilities returned by the planner to optimize the design of the environment itself. To do so, we propose integrating these motion planning algorithms into gradient-free optimization loops that operate over the design space of the robot’s environment. However, these algorithms are computationally intensive and thus will dominate the runtime of the optimization process. This motivates our second key insight: motion planners are not black boxes, but rather possess a unique structure which we can exploit to drastically increase the efficiency of our optimization. We dub this tight integration between the optimizer and motion planner a “white box” optimization approach, and propose two methods of actualizing it. Firstly, we leverage the lazy nature of many motion planners to restrict computations to only those relevant to the optimizer. Secondly, we frame evaluation of different robot environments within an online learning framework, which allows us to re-use information across subsequent utility queries.