Towards Integrating Hierarchical Goal Networks and Motion Planners to Support Planning for Human Robot Collaboration in Assembly Cells

Low-level motion planning techniques must be combined with high-level task planning formalisms to generate realistic plans that can be carried out by humans and robots. A representative example is planning for fenceless assembly cells where robots can collaborate seamlessly with humans to perform assembly tasks. Key constituent components include assembly sequence generation (Morato, Kaipa, and Gupta 2013), task decomposition between human and robot (Kaipa et al. 2014), system state monitoring (tracking human, robot, and assembly parts) (Morato et al. 2014a), instruction generation for humans (Kaipa et al. 2012), safety (Morato et al. 2014b), and error recovery (Morato et al. 2014a). In order to enable a coherent integration among these components, a high-level planner, interleaved with motion planners, is needed at several levels of the system hierarchy. For example, given a CAD model of a product to be assembled, motion planning methods can generate improved assembly precedence constraints (Morato, Kaipa, and Gupta 2013), which can be compiled into a high-level planning problem. Humans and robots share complimentary strengths. The planning framework can incorporate this knowledge to decompose the tasks effectively. Further, an integral planner must be able to perform plan-repair in order to handle contingencies: (1) low-level deviations in the geometric state without affecting the corresponding symbolic state (e.g., human places part in a wrong posture), which can be corrected at the motion planning level, or (2) deviations in the symbolic state (e.g., human picks incorrect part; improved alternative sequence may or may not exist), which needs to be corrected at both levels of planning. Task planning formalisms typically used to achieve this integration are Classical Planning (Cambon, Alami, and Gravot 2009; Erdem et al. 2011; Dornhege et al. 2009; Burbridge and Dearden 2013) and Hierarchical Task Network (HTN) Planning (Kaelbling and Lozano-Pérez 2011; Hadfield-Menell, Kaelbling, and Lozano-Perez 2013; Wolfe, Marthi, and Russell 2010). Whereas Classical planning is not scalable, HTNs impose stringent completeness requirements on domain models, which are difficult to guarantee in open, dynamic environments. Recently, we developed a new planning formalism called Hierarchical Goal Networks (HGNs) (Shivashankar et al. 2012; 2013) that combines scalability and expressivity advantages of HTNs and heuristic-search/reasoning capabilities of Classical planning into a single framework. In this work, we exploit the advantages of HGNs to tightly integrate it with motion planners. Our aims are twofold: (1) Design a generalpurpose planning-and-execution framework that combines HGN planning and execution-time plan-repair algorithms with off-the-shelf motion planners, and (2) Formulate this planning framework in the context of planning for human robot collaboration in assembly cells. The system takes as input the planning problem P (provides descriptions of the initial state, goals to be achieved, base action models at the task-planning level, control primitives at the motion planning level, and procedures to translate between symbolic and geometric state descriptions). P is input into an Offline Planning module, in which HGN planners and low-level motion planners interactively synthesize an executable plan structure Π that achieves the given goals when applied from the initial system state. Π is then input into an Execution-time Reasoner module to (1) monitor plan execution, and (2) repair Π in case the deviations from expected state render the current plan inexecutable.