Efficient Automatic Perception System Parameter Tuning On Site without Expert Supervision

Abstract

Many modern perception systems require human engineers to tune parameters in order to adapt to various environments and applications. This incurs a large startup cost when deploying a robotic system by relying on human expertise and ground truth instrumentation. To alleviate this, we propose a technique using empirical trials to automatically tune a perception system's parameters on-site without expert supervision. Our approach extends upon recent work on introspecting perception performance and uses Bayesian optimization to efficiently search the parameter configuration space. We validate our technique by tuning the laser and visual odometry systems of a physical ground robot in a variety of environments, achieving estimation errors competitive with baseline approaches that use ground truth.