Using Multiple Fidelity Models in Motion Planning

Abstract

Hospitals and warehouses use autonomous delivery robots to increase productivity. Robots must reliably navigate unstructured non-uniform environments which requires efficient long-term operation that robustly accounts for unforeseen circumstances. However, unreliable autonomous robots need continuous operator assistance, which decreases throughput and negates a robot's benefit. Planning with high fidelity models is more likely to lead to more robust plans, but is not needed in many situations. More specifically, a complex model is an inefficient use of plan-time computation resources when a robot navigates a flat simple environment, but a simple model, that can generate plans quickly, may fail to capture complex environment locales leading to task impediments.

This thesis presents a planning framework that reasons about multiple models for efficiently generating a single motion plan without sacrificing execution success. The framework chooses when to switch models and what model is most applicable within a single trajectory. This has the effect of focusing the use of complex models only when necessary. The framework also addresses uncertainty by adding variable uncertainty to models in the form of robot padding. The footprint padding is not fixed but is automatically chosen during plan-time based on the ability to find robust plans.

The approach is evaluated by simulating a mobile robot with attached trailer through various hospital environments. Our simulation experiments demonstrate that multi-fidelity model switching increases plan-time efficiency while still maintaining execution success.