Dynamic System Model Learning and Control with Barrier Methods: Applications to Robotics and Autonomous Systems Safety

New advancements in data-driven dynamic models and controller learning provide
new possibilities to enable skill transfer to robotic systems operating in various remote,
hazardous and confined e nvironments. This dissertation develops methods for
learning dynamic system models from the demonstration data obtained from its operations
in the paradigm of behavior cloning and related constrained controllers. The
behavior cloning methods can generate motion plans for robots to follow so they
can operate in a restricted domain. The method uses a dynamical system learning
approach for modeling the robot’s trajectory paths. In the proposed methods, the
robot’s end-effector motion trajectories are represented as solutions to the underlying
motion dynamics that can be learned using tools from machine learning. The model
parameters are learned such that the robot trajectories remain bounded in a prescribed
safe region and converge to the desired goal location with guarantees of the
model’s robustness to external disturbances. Constrained low-level controllers can be
used for the robot to track the reference trajectories or the motion plans generated by the dynamic system models. Thus, robot torque controllers for constrained systems
are developed, keeping the robot’s motion within joint limits. A mapping function
is first designed to transform the robot’s c onstrained dynamics to e quivalent unconstrained
ones. An adaptive controller is then developed to track the robot’s operation
within specified j oint l imits. The methods p resented i n t his d issertation a re tested
using experiments conducted on a 7-degrees-of-freedom dual-arm Baxter robot.