Author:

Abstract:

Still today, industrial robot assembly tasks are executed by replaying a preprogrammed trajectory, making them vulnerable to geometric uncertainties in the environment. Therefore, industrial assembly tasks are executed in expensive structured environments, which limits the use of robots to high volume and repetitive tasks, where the cost of the structured environment becomes relatively small. By equipping the robot with sensors, it can observe its environment and interact with its environment, allowing it to operate in less expensive, unstructured environments. Assembly tasks are only one example of compliant motion tasks, where a robot manipulates an object in contact with an environmental object. The force interaction between the contacting objects guides the manipulated object along the environmental object to overcome geometric uncertainties associated with the task. Aiming towards more intelligent and flexible robots, this thesis presents two high level approaches for sensor based compliant motion task specification. The first approach is based on a compliant motion path planner that generates a compliant path given the geometric models of the contacting objects. The path is expressed by the relative positions between the objects, and the desired contact formations. The second task specification method exploits the advanced manipulation skills of a human, to obtain a compliant path. While a human operator uses a demonstration tool to demonstrate the targeted compliant motion task, sensors on the demonstration tool measure contact forces, positions and velocities. Applying state of the art Bayesian sequential Monte Carlo methods (also known as particle filters), the sensor measurements are combined to simultaneously estimate continuous geometrical parameters and recognize the discrete contact formation between the objects. The simultaneous estimation is helped by the availability of a contact state graph of all possible contact formations. This thesis also presents the compliant task generator, an approach to automatically convert the output from the compliant path planner, or the output from a human demonstration, into a task specification for a hybrid robot controller. The planner or demonstration output is a geometric description of a compliant path, while the hybrid controller requires instantaneous force and velocity setpoints. The compliant task generator automatically converts a geometric compliant path into controller setpoints, allowing to plan or demonstrate a compliant motion task and immediately execute it on a real robot manipulator under active force control. Finally, this thesis uses the presented Bayesian estimators to monitor discrete contact formation transitions online, during the execution of a compliant motion task. This provides a feedback to the generator and controller components about transitions between different sub-tasks, and allows the system to select the most appropriate controller strategy for the current contact formation between the manipulated object and its environment.