Designing Certified Controllers to Prevent Falls for Legged Robots

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

This project will investigate controllers for legged robots that work reliably despite lacking full information about operating conditions, including uncertainty about terrain and robot properties. The approach differs from previous approaches in that the focus is entirely on maintaining balance. There are three integrated research tasks. First is creation of a feasible algorithm to determine starting positions of a legged robot from which falling can be avoided. Second is pre-computation of a so-called "safe set" of control inputs that is guaranteed to avoid falling. Both these tasks embed the hybrid, nonlinear dynamics of walking into a linear partial differential equation, which is propagated backwards in time from a set of acceptable final states to incorporate the effects of uncertainty. The third task is experimental evaluation of the results using a reconfigurable legged robot platform. As part of the third task, a human operator will be allowed to give arbitrary commands to a legged robot. A command filter will modify the user input so that it lies within the safe set. Thus the operator will be able to give any command without the robot falling. Legged robotic systems are an ideal candidate for search and rescue missions or nuclear power plant repair or disassembly. However successful teleoperation of such systems is challenging. Thus guaranteed safe operation will make these systems a viable means of keeping first responders out of harm's way. The techniques under study can be extended to other fields of robotics, such as active prostheses or exoskeletons, thereby benefitting the lives of those suffering from loss of mobility. This project will construct a novel optimization scheme for automated control synthesis to provide deterministic guarantees on the safety of a legged robotic system in the presence of model uncertainty. Rather than rely upon linearizations or assume arbitrary control authority, the approach considered in this project will address the following three aims: First, a new convex optimization tool to efficiently compute the set of states of a legged robotic system that have less than a user specified probability of avoiding falls in spite of uncertainty in the continuous and contact dynamics. Second, a novel control synthesis tool to pre-compute the set of controls that prevent falls for the set of states that can avoid falling. Finally, an interactive real-world setup wherein users can apply arbitrary inputs to a legged robot which are then modified to ensure safe operation across a variety of terrains; this setup will validate the robust synthesis method while allowing modelers to confidently explore configurations of a legged robot that have been rarely considered to date due to the limitations of existing control design techniques.