July 2009 Meeting:
Wednesday, July 15, 2009

Carnegie Mellon University, Silicon Valley (directions:   https://sv.cmu.edu/who_we_are/visitor)

Autonomous Control in Unknown Environments: Making Robots Inquisitive

Dr. Gabriel Hoffmann   

Autonomous vehicles are able to perform a broad variety of tasks, yet interacting with unknown environments largely remains a challenge. This talk examines the interaction between control actions and sensor observations. We discuss algorithms for robots that go beyond coping with uncertainty by being "inquisitive"--actively pursuing information to improve system performance-- and their application to autonomous helicopters and cars.

We first discuss the control system for an autonomous automobile from the DARPA Grand challenge, "Stanley", which traverses unknown environments and adapts its speed to maintain the ability to perceive terrain. This challenge of interacting sensing and control systems motivates the focus of the talk, that of controlling mobile sensors to acquire information.

To control mobile sensors, an information theoretic control algorithm is designed to enable the distributed, cooperative information-seeking. The algorithm aims to reduce the uncertainty of variables of interest at an optimal rate. By computing information theoretic terms directly from a probability distribution represented by a particle filter, the algorithm exploits a rich knowledge base build using past observations, with full knowledge of future sensing capabilities. Novel approximations allow large networks of vehicles to engage in cooperative sensing using decentralized control algorithms.

The mobile sensor control algorithms are demonstrated in simulations of three sensing modalities to perform an automated search for a lost target. These algorithms were experimentally implemented on STARMAC, a fleet of quadrotor helicopters. Flight experiments demonstrate autonomous search for an avalanche rescue beacon with one operator managing multiple autonomous, "inquisitive" helicopters.

Dr. Gabriel Hoffmann:

Gabriel Hoffmann is a researcher at the Palo Alto Research Center's Intelligent Systems Lab where he develops approaches for autonomous control and optimization of sensor-rich robots and embedded systems. His investigations explore the interaction between sensing systems, control systems, and physical systems.

Gabe received his Ph.D. in Aeronautics and Astronautics from Stanford in 2008. He developed algorithms to control mobile sensor networks to make optimal observations in their unknown surroundings. The methods combine information theory, probabilistic inference, and hybrid and optimal control. He designed quadrotor helicopter UAVs and used them to demonstrate cooperative autonomous search with a single person tasking multiple vehicles. While heading up control system development for the Stanford Racing Team, Gabe designed the control system that ran the autonomous car "Stanley", which won the DARPA Grand Challenge, and "Junior", which placed second in the Urban Challenge. He also researched fault tolerance for autonomous rendezvous for the NASA Exploration Program, and as an undergraduate at the UW Madison he led a group to design and fly experiments in reduced gravity through a NASA program.

https://hoffmann.stanford.edu