Combining task and motion planning for complex manipulation

Robots performing household tasks is a big challenge, where the operator/user needs to give sequential commands to the robot for each subtask. In order to perform a complicated task in say, a kitchen environment, task planners are needed that can reason over very large sets of states by manipulating partial descriptions, while geometric planners operate on completely detailed geometric specifications of world states. Home Exploring Robot Butler (HERB) currently uses a sequential task planner for table clearing which plans for the entire task and obtains a feasible global plan before beginning execution. We aim to implement a task planner that reduces the planning time by breaking the high level task into subgoals, making choices and committing to them, greatly reducing the length of plans. This approach is suitable for complex tasks where strict optimality is not crucial. In this report, we present the key features of HPN and its advantages that led us to use these key features (fluents, operators, etc.) in a sub-version that uses a breadth-first approach to plan in the now. INTRODUCTION Project Description Consider a robot unloading the dishwasher. Currently a user may have to specify a series of commands to make the robot carry out this task: drive to the dishwasher, open the dishwasher, pull out the top rack, pick up the cup, open the cupboard, put the cup into the cupboard, etc. We aim to enable HERB to take a high level task input, i.e. unload the dishwasher, and build a hierarchical plan to select, sequence and execute subtasks in the now. This project aims to use the ideas of HPN [1] to define a table clearing task and carry out interleaved planning and execution. The goal is to enable HERB to clear a table by picking up three cups placed in random configurations and placing them on a tray kept on the table. Motivation Robots in home environments need to deal with real time constraints. The main challenge we face here is the complex integration between a high-level task planner that selects the ordering of subtasks with low-level motion planners that generate arm motions to execute each subtask. Moreover, if the motion planning fails, for example, when the robot picks up the first two cups successfully but there is no motion plan to execute the subsequent tasks. Thus, there is a need for a robust task planner for a complicated sequence of tasks. RELATED WORK: HIERARCHICAL PLANNING IN THE NOW: (HPN) Hierarchical Planning in the Now is an approach to reduce planning time and plan length for carrying out a complex manipulation task. It operates on detailed, continuous geometric representations and does not require a-priori discretization of the state or action space. It is aggressively hierarchical. It makes choices and commits to them, exponentially decreasing the amount of search required. HPN can work effectively with non-determinism in the environment or in the low-level controllers where planning in detail far into the future will mostly be wasted, due to the inability to predict exactly what will happen. Planning ‘in the now’ will construct a plan at an abstract level, commit to it, and then recursively plan and execute actions to achieve the first step in the abstract plan without constructing the rest of the plan in detail. One of the risks associated with this approach is that the abstract plan might not be executable: the way that the first step is carried out could make it impossible to carry out subsequent steps, without undoing the results of earlier steps. This scenario is avoided by constraining the abstract plan to be serializable, i.e. for any step, there exist realizations for next steps. Interleaved planning and execution is a relatively standard regression-planning algorithm, based on an A* search that works backward from the goal, generating sub-goals that are the weakest precondition of the goal under each applicable action. It is used to solve single planning subproblems. The architecture can be thought of as doing a depth-first traversal of a planning tree, and is implemented as a recursive algorithm.