COAST: Constraints and Streams for Task and Motion Planning

TL;DR

COAST tackles the scalability problem in Task and Motion Planning by decoupling task planning from motion planning via a stream-based grounding step and constraint-driven feedback. It uses a plan-first, probabilistically-complete approach, with $P( ext{success}) o 1$ as the number of samples $n o fty$, that grounds task plans with stream objects after symbolic planning and then samples streams to verify feasibility. The method demonstrates order-of-magnitude reductions in cumulative task planning time across Blocks, Kitchen, and Rover domains compared with PDDLStream and IDTMP. This work broadens TAMP applicability to long-horizon robotic tasks and offers a practical, scalable framework.

Abstract

Task and Motion Planning (TAMP) algorithms solve long-horizon robotics tasks by integrating task planning with motion planning; the task planner proposes a sequence of actions towards a goal state and the motion planner verifies whether this action sequence is geometrically feasible for the robot. However, state-of-the-art TAMP algorithms do not scale well with the difficulty of the task and require an impractical amount of time to solve relatively small problems. We propose Constraints and Streams for Task and Motion Planning (COAST), a probabilistically-complete, sampling-based TAMP algorithm that combines stream-based motion planning with an efficient, constrained task planning strategy. We validate COAST on three challenging TAMP domains and demonstrate that our method outperforms baselines in terms of cumulative task planning time by an order of magnitude. You can find more supplementary materials on our project \href{https://branvu.github.io/coast.github.io}{website}.

Figures (4)

• Figure 1: We propose COAST, a sampling-based TAMP algorithm that is able to solve complex, geometrically constrained, long-horizon planning problems faster than prior state-of-the-art. We demonstrate the ability of our algorithm to solve problems from three domains: a $3 \times 3$ grid rearranging task (Blocksidtmp, left), a constrained pick-and-place kitchen task (Kitchenpddlstream, middle), a rover surveillance task with obstacles (Roverpddlstream, right).
• Figure 2: Percentage solved and cumulative task and motion planning times for the Blocks domain with increasing number of obstacles. On the most complex configuration (6 obstacles), our algorithm achieves 100% success while IDTMP achieves 20% and PDDLStream achieves 60%. The reported planning times include the failed trials that time out at 1200s. Our algorithm solves the largest problem two orders of magnitude faster than PDDLStream and IDTMP.
• Figure 3: Percentage solved and cumulative task and motion planning times for the Kitchen domain with increasing number of cook/clean goals. PDDLStream's planning process times out after 1200 seconds for 46% of the most challenging tasks (8 cook/clean goals), whereas our method achieves 100% success at magnitudes faster. PDDLStream's slow task planning times come from the explosive growth of its task state with stream objects, which our method avoids by introducing stream objects after task planning.
• Figure 4: Percentage solved and cumulative task and motion planning times for the Rover domain with increasing number of goal objects. We also include an ablation (purple) of our method with stream instance caching. PDDLStream has exponential growth in task planning time, whereas our task planning time remains nearly constant.

Theorems & Definitions (2)

• Theorem VII.1
• proof