VaPr: Variable-Precision Tensors to Accelerate Robot Motion Planning

Yu-Shun Hsiao, Siva Kumar Sastry Hari, Balakumar Sundaralingam, Jason Yik, Thierry Tambe, Charbel Sakr, Stephen W. Keckler, Vijay Janapa Reddi

Introduction

Modern robotics demands real-time motion planning for high-dimensional manipulators, often requiring collision-free trajectories to be generated in milliseconds. While GPU-accelerated planners like CuRobo have significantly reduced computation times, they have encountered a "memory wall." On edge devices such as the NVIDIA Jetson AGX Orin, the massive tensors required for trajectory optimization frequently exceed the L2 cache capacity. This results in constant data movement to and from off-chip DRAM, creating a performance bottleneck that limits the responsiveness of robotic systems.

To address this, we present VaPr (Variable-Precision), an optimization framework that identifies the minimum numerical precision required for individual data tensors. By strategically reducing bitwidths—in some cases down to just 4 bits—VaPr eliminates memory bottlenecks and achieves significant speedups without compromising the robot's ability to navigate complex environments safely.

The Challenge of Numerical Precision

A common approach to reducing memory overhead is global quantization, such as moving an entire pipeline from FP32 to FP16. However, motion planning is non-uniformly sensitive to numerical errors. While some parts of the optimization process can tolerate low precision, others—like the precise location of robot links—require high fidelity to avoid collisions. The challenge lies in finding a "Goldilocks" precision for every tensor: low enough to maximize memory efficiency, but high enough to maintain a 100% planning success rate.

Methodology

The VaPr framework optimizes the motion generation pipeline through a three-step process:

1. Characterization

The authors first identified five critical "bottleneck" tensors within the CuRobo pipeline. These tensors, primarily related to kinematics and collision checking, were found to be responsible for the majority of memory traffic. Analysis also revealed that several of these tensors (such as grad_out_spheres and closest_pt) are over 99% sparse, though the primary efficiency gains were found in bitwidth reduction rather than sparsity-aware computation alone.

2. Two-Phase Search Optimization

Because the search space for precision configurations is vast, VaPr employs a two-phase approach:

  • Phase 1 (Per-tensor Binary Search): The system determines the absolute minimum precision each tensor requires in isolation while maintaining all other tensors at FP32.
  • Phase 2 (Combinatorial Search): Since numerical errors propagate across kernels, the isolated minimums may not work in tandem. The framework uses an evolutionary algorithm (NSGA-II) to explore over 4 million possible combinations, identifying a set of precisions that work collectively to ensure planning success.

3. Implementation Paths

The researchers evaluated two implementation strategies:

  • Software-only: This version utilizes existing CUDA intrinsics to convert between FP32 and FP16, providing immediate benefits on current hardware.
  • Hardware-proposal: This version simulates specialized "FP converter units" capable of handling non-standard formats (e.g., FP4, FP6, FP8). This allows for even more aggressive bitwidth reduction and higher throughput.

Experimental Results

The framework was evaluated using the MotionBenchMaker dataset, focusing on a 7-DOF Franka Panda manipulator across eight diverse environments, including cluttered bookshelves and tabletop picking tasks. Testing was conducted on both the NVIDIA Jetson AGX Orin (edge) and the RTX 2080 Ti (desktop).

Memory and Speed Efficiency

VaPr successfully reduced the total bitwidth of the five targeted tensors from 160 bits (5 x 32-bit) to an average of just 43 bits—a nearly 4x reduction in memory footprint. This reduction translated directly into performance gains:

  • Software-only implementation: Achieved a 6.3% average speedup across both test platforms.
  • Hardware-accelerated implementation: Achieved speedups of 9.9% on the Jetson Orin and 17.7% on the RTX 2080 Ti.

Precision Sensitivity

A key insight from the study is that 4-bit floating-point formats are sufficient for many critical tensors in the motion generation stack. However, sensitivity varies by task; for instance, tensors representing the location of robot limbs (out_spheres) require higher precision (FP13-FP16) to prevent collisions, whereas tensors related to optimization costs can operate at much lower bitwidths.

Conclusion

VaPr demonstrates that high numerical precision is often "over-provisioned" in robotic motion planning. By adopting a variable-precision approach, developers can bypass the memory wall and achieve significant speedups on existing and future hardware. While the current implementation requires a brief environment-specific search phase and benefits most from proposed hardware units, it proves that "mixing and matching" precisions is a viable and necessary path forward for high-performance robotics. Future work will look toward expanding this technique to the entire motion planning stack and other high-dimensional tasks like rigid-body dynamics.

Summary updated

Abstract

High-dimensional motion generation requires numerical precision for smooth, collision-free solutions. Typically, double-precision or single-precision floating-point (FP) formats are utilized. Using these for big tensors imposes a strain on the memory bandwidth provided by the devices and alters the memory footprint, hence limiting their applicability to low-power edge devices needed for mobile robots. The uniform application of reduced precision can be advantageous but severely degrades solutions. Using decreased precision data types for important tensors, we propose to accelerate motion generation by removing memory bottlenecks. We propose variable-precision (VaPr) search optimization to determine the appropriate precision for large tensors from a vast search space of approximately 4 million unique combinations for FP data types across the tensors. To obtain the efficiency gains, we exploit existing platform support for an out-of-the-box GPU speedup and evaluate prospective precision converter units for GPU types that are not currently supported. Our experimental results on 800 planning problems for the Franka Panda robot on the MotionBenchmaker dataset across 8 environments show that a 4-bit FP format is sufficient for the largest set of tensors in the motion generation stack. With the software-only solution, VaPr achieves 6.3% and 6.3% speedups on average for a significant portion of motion generation over the SOTA solution (CuRobo) on Jetson Orin and RTX2080 Ti GPU, respectively, and 9.9%, 17.7% speedups with the FP converter.
BibTeX
@inproceedings{hsiao2023vapr,
  author = {Yu-Shun Hsiao and Siva Kumar Sastry Hari and Balakumar Sundaralingam and Jason Yik and Thierry Tambe and Charbel Sakr and Stephen W. Keckler and Vijay Janapa Reddi},
  title = {{VaPr: Variable-Precision Tensors to Accelerate Robot Motion Planning}},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems},
  year = {2023},
  doi = {10.1109/iros55552.2023.10342109},
}
Figure