Balancing actuation energy and computing energy in low-power motion planning

Abstract

Inspired by emerging low-power robotic vehicles, we identify a new class of motion planning problems in which the energy consumed by the computer while planning a path can be as large as the energy consumed by the actuators during the execution of the path. As a result, minimizing energy requires minimizing both actuation energy and computing energy since computing energy is no longer negligible. We propose the first algorithm to address this new class of motion planning problems, called Computing Energy Included Motion Planning (CEIMP). CEIMP operates similarly to other anytime planning algorithms, except it stops when it estimates that while further computing may save actuation energy by finding a shorter path, the additional computing energy spent to find that path will negate those savings. The algorithm formulates a stochastic shortest path problem based on Bayesian inference to estimate future actuation energy savings from homotopic class changes. We assess the trade-off between the computing energy required to continue sampling to potentially reduce the path length, the potential future actuation energy savings due to reduction in path length, and the overhead computing energy expenditure CEIMP introduces to decide when to stop computing. We evaluate CEIMP on realistic computational experiments involving 10 MIT building floor plans, and CEIMP outperforms the average baseline of using maximum computing resources. In one representative experiment on an embedded CPU (ARM Cortex A-15), for a simulated vehicle that uses one Watt to travel one meter per second, CEIMP saves 2.1-8.9x of the total energy on average across the 10 floor plans compared to the baseline, which translates to missions that can last equivalently longer on the same battery. As the the energy to move relative to the energy to compute decreases, the energy savings with CEIMP will increase, which highlights the advantage in spending computing energy to decide when to stop computing.