P3: Planning and Control

In mainstream robotics, task plans are almost exclusively imperative, that is, they consist of reci- pes of motion, sensing and decision-making actions to be realized by the robot. Those recipes are created at development time and are therefore inflexible and unable to deal with variations in objects and environments. The scientific challenge is develop declarative task plans. Declarative task planning means that, at development time, instead of motion plans, high-level task requirements are formulated as a combination of one or more objective functions to optimize and one or more constraints to satisfy. The robot must reason at runtime about how it will turn that declarative specification into a concrete recipe, given the current local and temporary con- text of the task execution.

Many harvesting and food processing tasks require two separate actuators (robot arms), e.g., when the product needs to be grasped first before it can be manipulated. Efficient processing requires a joint optimization of the overall plan and runtime coordination of the perception and control in both arms. Control approaches capable of dealing with the intricate and uncertain interaction mechanics during processing natural products will be devel- oped. As these interactions are virtually impossible to model analytically, reinforcement learning (learning from experience) will be used to optimize the controllers.

Project coherence: Planning and control (P3) will require active perception (P1) and a world model (P2). Con- versely, planning and control (P3) will provide guidance for active perception (P1). P3 also provides the plan- ning and control capabilities for use-case projects P5, P6 and P7.


Project partners: TUD-CR, TUe-CS

User involvement: Marel, Protonic, Maxon Motor, Houdijk, BluePrint Automation, Cerescon, ABB. See Chapter 8 for more details.