Autonomous System for Legged Robots: From Calibration and Pose Estimation to CLF Reactive Motion Planning

This dissertation seeks to develop a full autonomy system that allows bipedal robots to 1) acquire multi-modal data from a calibrated perception suite; 2) estimate their poses in textureless environments; 3) detect and avoid dynamic obstacles; 4) traverse unexplored, unstructured environments and undulating terrains; and 5) perform point-to-point topometric navigation. All the research presented in this dissertation focuses on advancing the state of the art in mobile robot autonomy. To ensure the practicality of our work, we have evaluated all of our algorithms on Cassie Blue. We develop an automatic pipeline for both LiDAR-camera extrinsic calibration and intrinsic calibration for LiDARs. The resulting calibrated system achieves pixel-level error when projecting a LiDAR point cloud on the camera’s image plane. Additionally, we propose a unifying intrinsic calibration technique for both spinning and solid-state LiDARs. Our method reduces by 48.7% the error in our factory-calibrated LiDAR. The calibrated sensors mounted on the perception suite enable us to open up an important chapter in Cassie's autonomy. Later, we develop the very first fiducial marker system for LiDAR point clouds --- LiDARTag --- to estimate a mobile robot's pose. Because LiDAR sensors are not affected by rapidly changing ambient lighting, LiDARTags can be used where cameras cannot, and the proposed fiducial marker can even operate in a completely dark environment. The LiDARTag system achieves millimeter translation error and a few degrees of rotation error, reaches 99.7% accuracy on decoding, and runs faster than 100 Hz. However, due to the quantization uncertainty and sparsity of LiDAR returns, the performance of the pose estimation when using a 32-beam Velodyne LiDAR is degraded when the target is farther than 12 meters. We therefore propose the concept of target shape optimization for enhancing pose and vertex estimation from LiDAR point clouds. Specifically, we design a target so that the edge points induced by LiDAR rings are ``highly'' sensitive to translation and rotation. We achieve centimeter error in translation and a few degrees of error in rotation even when an occluded target of width 0.96 meter is placed 30 meters away. Next, we propose an anytime iterative system to concurrently solve the multi-objective path planning problem and determine the visiting order of destinations. The system is comprised of an anytime informable multi-objective and multi-directional RRT* algorithm to form a simple connected graph, and a proposed solver that consists of an enhanced cheapest insertion algorithm and a genetic algorithm to solve the relaxed traveling salesman problem in polynomial time. Finally, we propose and experimentally demonstrate a reactive planning system for bipedal robots traversing unexplored, challenging terrains. The system consists of a low-frequency planning thread to find an asymptotically optimal path and a high-frequency reactive thread to accommodate robot deviation. The planning thread includes: a multi-layer local map to compute traversability for the robot on the terrain; an anytime omnidirectional Control Lyapunov Function for use with a Rapidly Exploring Random Tree Star that generates a vector field for specifying motion between nodes; a sub-goal finder when the final goal is outside of the current map; and a finite-state machine to handle high-level missions. The reactive thread copes with robot deviation while eliminating non-smooth motions via a vector field (defined by a closed-loop feedback policy) that provides real-time control commands to the robot's gait controller as a function of instantaneous robot pose.PHDRoboticsUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/174161/1/bjhuang_2.pdfhttp://deepblue.lib.umich.edu/bitstream/2027.42/174161/2/bjhuang_1.pd