Task and Motion Policy Synthesis for Mobile Manipulation

Robots should safely and correctly operate alongside uncontrollable agents such as humans. In domains involving uncontrollable agents, such as the scenario shown in Fig. 1, robots must react to changes online to ensure safety and accomplish desired tasks. Thus, rather than pre-programming finite instructions or pre-computing a single, linear plan, robots need a policy that determines the correct response over the infinite-horizon interaction with a changing environment, while also guaranteeing safety and achieving task goals. Task and Motion Synthesis (TMS) constructs such policy that satisfies both high-level task requirements and low-level motion constraints. In this abstract, we extend the TMS approach presented in [19] (henceforth referred as ROBOSYNTHREACT), to solve tasks with safety and recurrence goals (see Section II for definitions of safety and recurrence). At a high level, ROBOSYNTHREACT formulates TMS as a concurrent two-player game [6, 1] between the robot and the environment, and synthesizes a winning policy for the robot by iteratively generating a candidate and verifying its correctness [16]. This iterative policy synthesis procedure converges either to a winning policy or a proof that no such policy exists. We extend ROBOSYNTHREACT with two additional policy verification modules for safety and recurrence tasks. Like what we did in ROBOSYNTHREACT, we adapt the proof rules of [2] to develop these two policy verification modules, which provide compact, symbolic constraints that characterize the correctness of policies. Recently, there has been a growing interest in the integration of Task and Motion Planning (TMP) [9, 20, 3, 13, 10, 17, 4, 12]. These previous TMP approaches assume the environment is static, while we consider in this abstract a changing environment with uncontrollable agents where we require not just a plan describing a single execution path, but instead a policy describing the online response to uncontrollable events. Other recent work has focused on reactive synthesis for robots and hybrid systems [8, 21, 5, 18]. These approaches consider the differential dynamics of hybrid systems but do not incorporate efficient path planning. In contrast, we focus on the mobile manipulation domain where high dimensionality makes efficient, collision-free path planning crucial, and we therefore apply fast, sampling-based motion planning methods [14]. Furthermore, these previous works often perform a combinatorial search over large state spaces. In contrast, we avoid expensive combinatorial search using SMT-based symbolic Charge Region FoodPrep Region Chef Dishwasher Robot Countertop