Online Trajectory Optimization for Legged Robotics Incorporating Vision for Dynamically Efficient and Safe Footstep Locations

This paper presents a trajectory optimization algorithm for legged robotics that uses a novel cost function incorporating point cloud data to simultaneously optimize for footstep locations and center of mass trajectories. This novel formulation transforms the inherently discrete problem of selecting footstep locations into a continuous cost. The algorithm seamlessly balances the desire to choose footstep locations that enhance the dynamic performance of the robot while still choosing locations that are viable and safe. We demonstrate the success of this algorithm by navigating the ALPHRED V2 robotic system over unknown terrain in a simulation environment. INTRODUCTION Legged locomotion for robotic systems is an inherently underactuated problem requiring the complex planning of discrete foot placements and center of mass (CoM) trajectories in order to successfully navigate the world. Ground reaction forces at the points of contact and gravity are the only forces acting on a robot, making control of the ground reaction forces vital to achieve a successful motion plan. For this reason, foot placement is a critical component of developing a successful walking trajectory for a legged robot. However, due to the non-linear relationship between foot placement and robot dynamics many of the previous ∗Address all correspondence to this author. FIGURE 1. SIMULATION OF ALPHRED V2 TRAVERSING