Interactive search for action and motion planning with dynamics

This paper proposes an interactive search approach, termed INTERACT, which couples sampling-based motion planning with action planning in order to effectively solve the combined task and motion planning problem. INTERACT is geared towards scenarios involving a mobile robot operating in a fully known environment consisting of static and movable objects. INTERACT makes it possible to specify a task in the planning domain definition language (PDDL) and automatically computes a collision-free and dynamically feasible trajectory that enables the robot to accomplish the task. The coupling of sampling-based motion planning with action planning is made possible by expanding a tree of feasible motions and partitioning it into equivalence classes based on the task predicates. Action plans provide guidance as to which a equivalence class should be further expanded. Information gathered during the motion tree expansion is used to adjust the action costs in order to effectively guide the expansion towards the goal. This interactive process of selecting an equivalence class, expanding the motion tree to implement its action plan and updating the action costs and plans to reflect the progress made is repeated until a solution is found. Experimental validation is provided in simulation using a robotic vehicle to accomplish sophisticated pick-and-place tasks. Comparisons to previous work show significant improvements.

本文提出一种名为INTERACT的交互式搜索方法，该方法将基于采样的运动规划与动作规划相结合，以有效求解任务与运动规划联合问题。INTERACT适用于移动机器人在包含静态与可移动物体的完全已知环境中作业的场景。该方法支持通过规划域定义语言（PDDL）指定任务，并可自动计算出无碰撞且动态可行的轨迹，使机器人得以完成指定任务。通过扩展可行运动树并基于任务谓词将其划分为若干等价类，可实现基于采样的运动规划与动作规划的耦合。动作规划将提供指导，指明应进一步扩展哪一个等价类。在运动树扩展过程中收集到的信息将用于调整动作代价，以有效引导扩展过程朝向目标推进。这一选择等价类、扩展运动树以执行其动作规划、并更新动作代价与规划以反映当前进展的交互式过程将持续重复，直至找到可行解。本文通过仿真实验，使用移动机器人完成复杂拾取-放置任务对所提方法进行了验证。与现有相关工作的对比结果表明，该方法性能提升显著。