| Summary: | Contact-rich manipulation tasks, such as assembly, require a robot to reason about both the geometric relationship between parts as well as the dynamical relationship between the forces the robot exerts and the motion of the parts. The application of forces enables the robot to reduce its uncertainty by purposefully contacting the environment, a crucial skill in real-world domains where state is not fully observed. In this thesis, a planner is introduced that reasons over both gripper poses and joint stiffnesses, trading off motion generation to reach an objective and force production to manage uncertainty. Our planner performs a greedy optimization over stiffness and learns a model of the relationship between control output and goal achievement to bias the pose search. This planner is validated on a peg-in-hole insertion task in simulation and the real world and a puzzle assembly task in simulation. We measure the effects of solving for stiffnesses and generating robust gripper poses in terms of the uncertainty our planner can address.
|
|---|