A Modular Architecture for Office Delivery Robots
Reid Simmons
Richard Goodwin, Karen Zita Haigh, Sven Koenig, Joseph O'Sullivan
School of Computer Science, Carnegie Mellon University
Office delivery robots have to perform many tasks. They have to
determine which office to visit next, plan a path to that office,
follow that path reliably, and avoid obstacles in the process. We
present a complete robot architecture that addresses all of these
issues in novel ways. Our architecture accounts for noisy robot
sensors and actuators, a dynamic and partially unknown environment,
and real-time behavior despite limited processing power. The
architecture is composed of four abstraction levels: Obstacle
avoidance is performed by our Curvature-Velocity Method; Navigation is
done using Partially Observable Markov Decision Process Models; Path
Planning uses a decision-theoretic generate, evaluate and refine
strategy with sensitivity analysis; and Task scheduling is performed
using a symbolic planning architecture. The levels are implemented as
independent processes that use the Task Control Architecture to
communicate among each other. A version of our robot architecture has
been in daily use since December 1995. Since then, our robot has
served over 1500 navigation requests that were specified using our
World Wide Web and Zephyr interfaces, and traveled a total of more
than 60 kilometers in the process.