pRRTC: GPU-Parallel RRT-Connect for Fast, Consistent, and Low-Cost Motion Planning

TL;DR

The paper tackles the high computational cost of sampling-based motion planning in high-dimensional robots by introducing pRRTC, a GPU-accelerated RRT-Connect planner co-designed for multi-level parallelism. It combines concurrent tree expansion, SIMT-optimized collision checking, and efficient memory management to accelerate planning across iterations and primitive operations. Across 7, 8, and 14 DoF benchmarks on MotionBenchMaker, pRRTC delivers up to 10× speedups, 5.4× reduction in planning-time standard deviation, and 1.4× lower initial path cost, while enabling real-time planning with dynamic obstacles on a 14-DoF dual-arm setup and open-sourcing the implementation. The work demonstrates that software-hardware co-design on GPUs can dramatically improve SBMP performance, offering practical impact for real-time robotic planning in cluttered or dynamic environments.

Abstract

Sampling-based motion planning algorithms, like the Rapidly-Exploring Random Tree (RRT) and its widely used variant, RRT-Connect, provide efficient solutions for high-dimensional planning problems faced by real-world robots. However, these methods remain computationally intensive, particularly in complex environments that require many collision checks. To improve performance, recent efforts have explored parallelizing specific components of RRT such as collision checking, or running multiple planners independently. However, little has been done to develop an integrated parallelism approach, co-designed for large-scale parallelism. In this work we present pRRTC, a RRT-Connect based planner co-designed for GPU acceleration across the entire algorithm through parallel expansion and SIMT-optimized collision checking. We evaluate the effectiveness of pRRTC on the MotionBenchMaker dataset using robots with 7, 8, and 14 degrees of freedom (DoF). Compared to the state-of-the-art, pRRTC achieves as much as a 10x speedup on constrained reaching tasks with a 5.4x reduction in standard deviation. pRRTC also achieves a 1.4x reduction in average initial path cost. Finally, we deploy pRRTC on a 14-DoF dual Franka Panda arm setup and demonstrate real-time, collision-free motion planning with dynamic obstacles. We open-source our planner to support the wider community.

Figures (7)

• Figure 1: A conceptual illustration of pRRTC in an abstract configuration space operating across three levels of GPU-parallelism. a) Separate GPU blocks extend both the start and goal trees with multiple samples in parallel. b) Within each block, parallel groups of four threads compute collision checks across the multiple configurations that correspond intermediate states of a motion. c) Individual threads parallelize the underlying kinematic linear algebra needed for these collision checks. Individual threads also parallelize nearest neighbors search.
• Figure 2: The MotionBenchMaker scenes provide diverse manipulation problems for 7, 8, and 14 DoF robots. The problems include counter top manipulation, accessing shelves, and constrained reaching. Panda and Fetch have 7 scenes with 100 problems each, while Baxter has 3 scenes with 500 problems each.
• Figure 3: Empirical Cumulative Distribution Functions (ECDFs) for all Software Benchmarks.
• Figure 4: MotionBenchMaker planning time and cost by problem on Fetch (8 DoF). Problems are ordered from left to right by increasing complexity. All times are shown on a logarithmic scale. pRRTC achieves a 5× speedup in average planning time, 10× speedup on the computationally challenging Cage problems, and 8.4× decrease in planning time standard deviation.
• Figure 5: MotionBenchMaker planning time across different ablations of pRRTC on Baxter (14 DoF). All times are shown on a logarithmic scale. Each one of the design choices individually improved the efficiency of pRRTC, with their combined effects achieving a 41× speedup over a naive parallel baseline.