Resumen de Physics-based motion planning for grasping and manipulation
This thesis develops a series of knowledge-oriented physics-based motion planning algorithms for grasping and manipulation in cluttered an uncertain environments. The main idea is to use high-level knowledge-based reasoning to define the manipulation constraints that define the way how robot should interact with the objects in the environments. These interactions are modeled by incorporating the physics-based model of rigid body dynamics in planning.
The first part of the thesis is focused on the techniques to integrate the knowledge with physics-based motion planning. The knowledge is represented in terms of ontologies, a prologbased knowledge inference process is introduced that defines the manipulation constraints. These constraints are used in the state validation procedure of sampling-based kinodynamic motion planners. The state propagator of the motion planner is replaced by a physics-engine that takes care of the kinodynamic and physics-based constraints (Muhayyuddin et al., 2015;Muhayyuddin et al., 2016). To make the interaction human-like, a low-level physics-based reasoning process is introduced that dynamically vary the control bounds by evaluating the physical properties of the objects. As a result, power efficient motion plan are obtained (Muhayyuddin et al., 2017a). Furthermore, a framework has been presented to incorporate linear temporal logic within physics-based motion planning to handle complex temporal goals (Muhayyuddin et al., 2017b).
The second part of this thesis develops physics-based motion planning approaches to plan in cluttered and uncertain environments (Muhayyuddin et al., 2018a; Muhayyuddin et al., 2018b). The uncertainty is considered in 1) objects’ poses due to sensing and due to complex robot-object or object-object interactions; 2) uncertainty in the contact dynamics (such as friction coefficient); 3) uncertainty in robot controls. The solution is framed with samplingbased kinodynamic motion planners that solve the problem in open-loop, i.e., it considers uncertainty while planning and computes the solution in such a way that it successfully move the robot from start to the goal configuration even if there is uncertainty in the system.
To implement the above stated approaches, a knowledge-oriented physics-based motion planning tool is presented (Muhayyuddin et al., 2017c). It is developed by extending The Kautham Project, a C++ based tool for sampling-based motion planning. Finally, the current research challenges and future research directions to extend the above stated approaches are discussed.
