Robust Manipulation with Active Compliance

Human manipulation skills are filled with creative use of physical contacts to move the object about the hand and in the environment. However, it is difficult for robot manipulators to enjoy this dexterity since contacts may cause the manipulation task to fail by introducing huge forces or unexpected change of constraints, especially when modeling uncertainties and disturbances are present. A properly designed robot compliance can provide the robot with the resilience and reliability in handling contacts.

This thesis proposes a framework for robust manipulation with contacts using active compliance. We provide quasi-static modeling that shows the necessity of compliance in rigid body manipulation. We further identify two causes of failure in manipulation: kinematic ill-conditioning and unexpected change of contact modes, and illustrate how the robot compliance can help avoid those failure types in manipulation tasks. First, we propose robustness metrics for each type of the failures. The metrics measure the amount of modeling uncertainty and the magnitude of external disturbance forces the system can take before a failure happens. Second, we provide methods to optimize the robustness metrics in a compliance control setting, as well as methods to improve the robustness of a contact-implicit

motion planning. Finally, we experimentally validate our proposed approaches in a variety of manipulation

problems. Our method efficiently finds solutions with consistent high quality during testing. The result shows that our framework trades-off well between model complexity and accuracy, captures major factors in manipulation problems

while keeping a low computation burden.