Modeling, System Identification, and Control for Dynamic Locomotion of the LittleDog Robot on Rough Terrain

In this thesis, I present a framework for achieving a stable bounding gait on the LittleDog robot over rough terrain. The framework relies on an accurate planar model of the dynamics, which I assembled from a model of the motors, a rigid body model, and a novel physically-inspired ground interaction model, and then identified using a series of physical measurements and experiments. I then used the RG-RRT algorithm on the model to generate bounding trajectories of LittleDog over a number of sets of rough terrain in simulation. Despite significant research in the field, there has been little success in combining motion planning and feedback control for a problem that is as kinematically and dynamically challenging as LittleDog. I have constructed a controller based on transverse linearization and used it to stabilize the planned LittleDog trajectories in simulation. The resulting controller reliably stabilized the planned bounding motions and was relatively robust to significant amounts of time delays in estimation, process and estimation noise, as well as small model errors. In order to estimate the state of the system in real time, I modified the EKF algorithm to compensate for varying delays between the sensors. The EKF-based filter works reasonably well, but when combined with feedback control, simulated delays, and the model it produces unstable behavior, which I was not able to correct. However, the close loop simulation closely resembles the behavior of the control and estimation on the real robot, including the failure modes, which suggests that improving the feedback loop might result in bounding on the real LittleDog. The control framework and many of the methods developed in this thesis are applicable to other walking systems, particularly when operating in the underactuated regime. Thesis Supervisor: Russ L. Tedrake Title: Associate Professor of Computer Science and Engineering