Combining learning from demonstration and search algorithm for dynamic goal-directed assembly task planning

In this paper, a learning approach is proposed to enable robots to generate assembly plans to assist users during assembly tasks. This plan is generated using a CAD model that represents a fully assembled goal object. The CAD model is induced by tracking a demonstrator assembling the components of that object. Forward assembly planning is an NP-hard problem, but we introduce pruning methods for the search tree that make the approach practical. Our dynamic planner generates an assembly plan that a user can follow to reproduce an identical object. Our system guides the user during the assembly task by suggesting parts to connect and how to connect them. If the user deviates from the suggested plan, the system analyses the unexpected state and determines whether the user’s action has brought the partially assembled object to a doomed state from which the goal state is not reachable. In this case, the user is warned about the dead-end. Otherwise, the system dynamically revises the current assembly plan. The system is validated with experiments on IKEA and LEGO objects.