Feasibility and Optimization of Fast Quadruped Walking with One- Versus Two-at-a-Time Swing Leg Motions for RoboSimian

This paper presents two planning methods for generating fast walking gaits for the quadruped robot RoboSimian and analyzes their feasibility and performance as a function of joint velocity limits. One approach uses a two-at-a-time swing leg motion gait, generated using preview control of the zero moment point (ZMP) on a narrow, double-support base of support. The second method uses a one-at-a-time swing leg crawl gait. We employ and compare both value iteration and gradient descent methods to generate the motions of the swing leg through free space for the crawl gait. For this gait, our optimization strategies exploit redundant degrees of freedom in the limb, which proves to increase walking speed significantly. By contrast, the ZMP-based gait requires nearly symmetric motions of the (fixed) stance legs and (free) swing legs with respect to the body, resulting in limited ability to improve the joint trajectories of the limbs through clever planning, and so an inverse kinematics (IK) table solution is used for all limb kinematics here. Our ZMP-based planning is only feasible if joint velocity limits are sufficiently large. Also, both the overall step size and speed of this gait vary nonlinearly as velocity limits increase. Lastly, comparing the two methods, we show that walking speed is faster using two-at-a-time swing leg motions when the joint velocity limit is between 1.62 and 3.62 rad/sec and that the crawl gait is optimal otherwise.