High-Performance Dual-Arm Task and Motion Planning for Tabletop Rearrangement

TL;DR

SDAR tackles dual-arm task and motion planning for dense tabletop rearrangement with entangled object dependencies. It combines a dependency-graph driven task planner with a GPU-accelerated, sampling-based motion planner, plus a robust failure-recovery mechanism. The key contributions are dependency-driven dual-arm task decomposition, synchronized motion generation with untangling, and near real-time performance with 100% simulated success and hardware transfer to UR5e. The results show significantly higher success and lower execution time than previous state-of-the-art baselines, highlighting practical potential for efficient dual-arm manipulation in cluttered settings.

Abstract

We propose Synchronous Dual-Arm Rearrangement Planner (SDAR), a task and motion planning (TAMP) framework for tabletop rearrangement, where two robot arms equipped with 2-finger grippers must work together in close proximity to rearrange objects whose start and goal configurations are strongly entangled. To tackle such challenges, SDAR tightly knit together its dependency-driven task planner (SDAR-T) and synchronous dual-arm motion planner (SDAR-M), to intelligently sift through a large number of possible task and motion plans. Specifically, SDAR-T applies a simple yet effective strategy to decompose the global object dependency graph induced by the rearrangement task, to produce more optimal dual-arm task plans than solutions derived from optimal task plans for a single arm. Leveraging state-of-the-art GPU SIMD-based motion planning tools, SDAR-M employs a layered motion planning strategy to sift through many task plans for the best synchronous dual-arm motion plan while ensuring high levels of success rate. Comprehensive evaluation demonstrates that SDAR delivers a 100% success rate in solving complex, non-monotone, long-horizon tabletop rearrangement tasks with solution quality far exceeding the previous state-of-the-art. Experiments on two UR-5e arms further confirm SDAR directly and reliably transfers to robot hardware.

Figures (10)

• Figure 1: Illustration of the dual-arm tabletop rearrangement setup examined in this work. (a) Motion planner performing regular pick-and-place, (b) Arms working in close proximity to handle objects that are close to each other, (c) Falling back to single-arm sequential execution when no feasible dual-arm solution can be quickly found.
• Figure 2: Illustration of a tabletop rearrangement instance that will be used as a running example. $\mathcal{W}$ is the workspace (a) Start configuration $O^s$. (b) Goal configuration $O^g$. (c) The induced dependency graph $\mathcal{G}_e$.
• Figure 3: Illustration of possible sampled grasp poses for grasping the two cuboids. For the arm on the right, the grasp poses are the same as the top-down ones. For the left arm, each is a different pose from a different approaching angle.
• Figure 4: Examples of start and goal configurations along with their corresponding dependency graphs. In the goal panels, the start configurations are also shown as white squares for reference. The four categories are shown: (top left) random configuration, (top right) single-cycle configuration, (bottom left) double-cycle configuration, and (bottom right) mixed configuration containing independent tasks, chains, and cycles. Black arrows denote chain dependencies in the task graph, while red and blue arrows represent different cycles.
• Figure 5: Execution time for all simulation-evaluated test cases. Where the horizontal legends are as defined at the start of this section (e.g., R10 means the random setting with 10 objects). Missing bars for a given test case and algorithm combination indicate the method failed to produce a valid solution.