Enabling Autonomous Legged Robot Agility
Successful deployment of legged robots in industrial applications such as environmental monitoring, material handling, or inspection requires that platforms perform complicated tasks quickly in unstructured environments. Unlocking the capabilities offered by legged morphologies requires autonomously harnessing their capability for agility to perform these tasks. However, autonomous and agile legged robot locomotion over unstructured terrain remains difficult due to underactuation, the hybrid nature of intermittent contact, as well as kinematic, dynamic, and computational constraints. Overcoming these restrictions requires some level of reasoning about how best to navigate the environment, yet often it is not clear what level of reasoning is best to solve a task. Simple models enable efficient computation and are generally robust to many forms of error, but they cannot capture all the salient features of a complex, multi-behavioral system such as a legged robot. Highly detailed models may be able to capture these features, but often at the expense of physical and computational robustness.
This thesis seeks to overcome this apparent trade-off and enable autonomous legged robot agility by adapting the complexity of the solution to the task at hand. These methods include a global motion planner with mixed-complexity motion primitives and an accompanying opensource full stack software framework for deployment on agile quadrupeds, a novel form of model predictive control that adapts its model to ensure feasibility and improve efficiency, and developments in bio-inspired robotic tail design to improve stability while reducing overhead costs.
The novel global motion planner enables autonomous agility by constructing long-horizon plans in real-time. It employs a mixture of motion primitives of varying fidelity to navigate from the current position to the goal while avoiding (or leaping over) obstacles and uneven terrain. This planner is shown to compute nearly an order of magnitude faster than comparable algorithms over dynamically challenging environments.
This global planner is featured within an open-source software package – Quad-SDK – which enables other researchers to deploy agile autonomy to their quadrupedal platforms. The package implements a full planning and control hierarchy which also includes a nonlinear model predictive controller with a novel warm starting technique that enables highly stable execution of dynamic motion plans. Several experiments are shown in simulation and hardware which demonstrate the system’s ability both plan and execute long horizon motion plans which include leaps. The package is also well supported with software tools to enable rapid development.
Even with efficient global planners, employing reduced-order models will generally result in locomotion which encounters constraints in the full-order representation of the system. This work presents a form of model predictive control which adapts the complexity of the model to capture the salient dynamics and constraints of the task. It is shown that under certain well-known template and anchor conditions this method yields provable stability properties, and enables simplification of the optimal control problem which results in improved performance. This method is benchmarked against fixed complexity formulations over candidate dynamic behaviors and is shown to be more stable than reduced-order configurations and more efficient than full-order configurations.
The novel tail design explores aerodynamic drag as a tool to regulate angular momentum independent of foot contact for improved agility. A new metric for the effectiveness of aerodynamic drag is presented and analyzed. Comparison to standard inertial effectiveness shows that aerodynamic drag tails can perform the same reorientation tasks as inertial tails but for a fraction of the inertia. The utility of these tails is demonstrated in hardware for disturbance rejection and locomotion assistance tasks, showcasing their enhanced practicality over inertial tails in providing leg-independent momentum regulation.
- Mechanical Engineering
- Doctor of Philosophy (PhD)