A methodology to quantify risk of failure for dynamic robots

Abstract

Humans possess an innate sense of danger that guides the execution of extraordinary dynamic maneuvers. They can also use this sense to generate creative recovery strategies to eventually come to a safe stop. This capability is not yet available to robots, fundamentally because there is no clear metric that represents the quantified risk of failing. Possessing such a metric would allow robots to explore their dynamic capability up to their physical limitations. This thesis attempts to address this problem by introducing a methodology to quantify the risk of failure for dynamic robots. It employs a sampling-based network constructed using the principles of viability theory, which focuses on the avoidance of failure instead of the regulation to specific movements. Simplifications that specifically target complex hybrid systems are explored to extend the usage of viability theory for practical application to legged robots. The results of this methodology are the Viable State Network, a network showing the non-failing and failing states, and the Risk Map, the quantified risk of failure. These concepts are demonstrated for a planar hopping robot model.