First Steps toward Automatically Generating Bipedal Robotic Walking from Human Data

This paper presents the first steps toward automatically generating robotic walking from human walking data through the use of human-inspired control. By considering experimental human walking data, we discover that certain outputs of the human, computed from the kinematics, display the same “universal” behavior; moreover, these outputs can be described by a remarkably simple class of functions, termed

canonical human walking functions

, with a high degree of accuracy. Utilizing these functions, we consider a 2D bipedal robot with knees, and we construct a control law that drives the outputs of the robot to the outputs of the human. Explicit conditions are derived on the parameters of the canonical human walking functions that guarantee that the zero dynamics surface is partially invariant through impact, i.e., conditions that guarantee

partial hybrid zero dynamics

. These conditions therefore can be used as constraints in an optimization problem that minimizes the distance between the human data and the output of the robot. In addition, we demonstrate through simulation that these conditions automatically generate a stable periodic orbit for which the fixed point can be explicitly computed. Therefore, using only human data, we are able to automatically generate a stable walking gait for a bipedal robot which is as “human-like” as possible.