This thesis presents a novel learning-based framework designed to enhance the efficiency and scalability of Task and Motion Planning (TAMP) for robotic manipulation tasks within cluttered and intricate environments. Conventional TAMP approaches encounter significant hurdles arising from the high computational burden of repetitive geometric feasibility assessments and the exponential growth of the search space, particularly in multi-robot scenarios. In response to these challenges, the thesis puts forth two deep learning models: AGFPNet, which predicts the feasibility of manipulation actions to minimize reliance on geometric planners, and GRN, which employs graph neural networks and scene graphs to offer interpretable and adaptable feasibility predictions, including insights into the causes of action failure.
Furthermore, these models are expanded to accommodate collaborative multi-robot scenarios and intricate, mesh-shaped objects. Leveraging these models, the thesis suggests a multi-robot TAMP algorithm guided by feasibility insights, which seamlessly integrates learned geometric reasoning into a search-based planner. This integration facilitates expedited search space exploration by prioritizing promising actions while avoiding unfeasible ones. Additionally, a new informed backtracking mechanism is introduced, utilizing predicted infeasibility causes to refine planning constraints and heuristic cost functions towards the objective.
Extensive experiments showcased the efficacy of the proposed approach in accelerating planning processes, enhancing scalability, and preserving generalizability, even when operating in highly cluttered or expansive environments. The outcomes of these experiments demonstrated the ability of the proposed method to successfully tackle complex manipulation tasks with extended planning horizons in challenging, cluttered environments, while significantly reducing planning times.
About the Author
