Global Tensor Motion Planning

An T. Le, Kay Hansel, Joao Carvalho, Joe Watson, Julen Urain, Armin Biess, Georgia Chalvatzaki, and Jan Peters

Motion planning with probabilistic completeness has been a foundation of robotics research, with seminal works like PRM and RRTConnect serving as foundational methods for years. However, as the complexity of robotic tasks grows, there is a rising demand for batch-planning methods. Several factors drive this interest: (i) the need to gather large datasets for policy learning, (ii) the inherent non-linearity of task objectives that lead to multiple local minima, and (iii) the increasing availability of powerful GPUs or TPUs. Despite these advancements, batching traditional path planning algorithms, such as RRT/PRM and their variants, remains an ongoing challenge. This paper revisits classical motion planning, introducing a simple discretation structure with layers of waypoints, which can be represented as tensors, enabling GPU/TPU utilization. We propose Global Tensor Motion Planning (GTMP), which enables highly batch-able operations on multiple planning instances, such as batch collision checking and batch value iteration, while maintaining an easily vectorizable implementation with JAX. This simplicity allows differentiable planning and rapid integration with modern frameworks, making the algorithm particularly desirable for real-time applications and scalable robot learning.

I. Lidar-scanned Occupancy Demos

In these examples, we vectorize planning with 100 paths. It took ~0.2 ms on an RTX3090 GPU. The parameters are (M = 2, N = 200, H = 100). The maps are ACES3 Austin, Intel Lab, and Orebro. The first and second rows demonstrate the solutions of GTMP and GTMP-Akima, respectively. GTMP-Akima does not perform very well in narrow passage problems.

II. Motion Bench Maker Dataset Panda Demos

In these examples, we vectorize planning with 100 paths. It took ~0.2 ms on an RTX3090 GPU. The parameters are (M = 1, N = 50, H = 20). All GIFs show the successful paths consecutively.