DRACO: Co-design for DSP-Efficient Rigid Body Dynamics Accelerator

Xingyu Liu, Jiawei Liang, Yipu Zhang, Linfeng Du, Chaofang Ma, Hui Yu, Jiang Xu, Wei ZhangarXiv 2025 · DOI · arXiv

Abstract

Rigid Body Dynamics (RBD) computation is the fundamental engine of modern robotics, yet it remains a critical bottleneck in advanced control schemes like Model Predictive Control (MPC), where it can consume up to 90% of total processing time. While FPGAs offer significant potential for low-latency parallel processing, existing accelerators are frequently limited by the "DSP wall," high-latency reciprocal operations, and inefficient resource utilization. We present DRACO, a novel FPGA-based accelerator that overcomes these barriers through a comprehensive hardware-algorithm co-design. By integrating a precision-aware quantization framework, a division-deferring algorithm for mass matrix inversion, and dynamic inter-module DSP reuse, DRACO achieves up to an 8× throughput increase and 7.4× latency reduction over state-of-the-art solutions. Crucially, DRACO enables 1kHz real-time control for high-degree-of-freedom (DOF) robots, such as the 29-DOF Atlas humanoid, which was previously unattainable for complex MPC horizons.

Introduction

The computational complexity of Rigid Body Dynamics (RBD) scales aggressively with a robot's degrees of freedom (DOF). As the industry moves toward complex systems like 12-DOF quadrupeds and 29-DOF humanoids, traditional CPU and GPU solutions struggle to meet the strict latency requirements of real-time control. FPGAs are theoretically ideal for this task, but they face three major roadblocks:

  1. The DSP Wall: High-DOF robots require massive parallel multiply-accumulate (MAC) operations that exhaust the FPGA's Digital Signal Processing (DSP) slices.
  2. The Reciprocal Bottleneck: Calculating the mass matrix inverse ($M^{-1}$) involves serial reciprocal operations that create significant latency on the critical path.
  3. Rate Mismatches: Imbalances in processing speeds between different mathematical modules lead to idle hardware cycles and low overall efficiency.
Motivation for DRACO: High-DOF robots increase MAC operations, leading to control rates below the 1kHz best-practice threshold on standard hardware.

Methodology: The DRACO Architecture

DRACO introduces three key innovations to address the identified bottlenecks:

Precision-Aware Quantization Framework

Unlike standard 32-bit floating-point approaches, DRACO utilizes the inherent physical robustness of robotic systems. Using an Iterative Control and Motion Simulator (ICMS), the authors determined that robots can maintain sub-millimeter trajectory accuracy with bit-widths as low as 18 or 24 bits. This precision-aware quantization reduces DSP demand by up to 4×, allowing more complex models to fit on a single FPGA.

Division Deferring for $M^{-1}$ Computation

Reciprocal operations are notoriously slow on FPGAs. DRACO reformulates the mass matrix inversion algorithm to "defer" these divisions. By implementing a "holding division" mechanism, the reciprocal operations are removed from the longest latency path (the backward pass) and executed in parallel with other matrix operations. This algorithmic reordering hides the division latency, significantly boosting throughput.

Inter-Module DSP Reuse

To maximize hardware efficiency, DRACO employs a dynamic resource-sharing architecture. In typical pipelines, modules like Inverse Dynamics (ID) and Mass Matrix Inversion (Minv) may have mismatched processing rates, leading to idle DSP slices. DRACO's architecture allows these resources to be shared across modules; when one module is idle or waiting for data, its DSP slices are reallocated to active modules, ensuring constant high-throughput processing.

Overview of Challenges and Contributions: DRACO maps specific hardware bottlenecks to algorithmic and architectural solutions.

Evaluation and Results

The researchers evaluated DRACO using standard URDF models for the KUKA iiwa (7-DOF), HyQ (12-DOF), and Atlas (29-DOF). Performance was compared against the Pinocchio (CPU) and GRiD (GPU) libraries, as well as the Roboshape and Dadu-RBD FPGA accelerators.

Key Performance Metrics

  • Throughput and Latency: DRACO achieved an 8× throughput boost and a 7.4× reduction in latency compared to Dadu-RBD.
  • Real-Time Feasibility: For the 29-DOF Atlas humanoid, DRACO successfully maintained a 1kHz control rate for MPC with long-horizon settings, a critical benchmark for stable humanoid locomotion.
  • Resource Efficiency: The quantization framework enabled a 4× reduction in DSP usage without compromising control stability.
  • Power Consumption: DRACO demonstrated high efficiency, consuming approximately 33.5W for the KUKA iiwa model, comparable to previous designs but with vastly superior performance.

Discussion and Conclusion

DRACO demonstrates that the computational limits of high-DOF robotics can be overcome by moving beyond generic hardware acceleration toward a co-designed approach. By exploiting the physical properties of robots—such as error tolerance—and reordering mathematical operations to suit FPGA strengths, DRACO sets a new standard for real-time robotic control. Future research will investigate mixed-precision schemes to further optimize specific RBD components and explore the potential for hardware-agnostic ASIC implementations to drive even greater power efficiency in edge robotics.

Summary updated

Abstract

We propose a hardware-efficient RBD accelerator based on FPGA, introducing three key innovations. First, we propose a precision-aware quantization framework that reduces DSP demand while preserving motion accuracy. This is also the first study to systematically evaluate quantization impact on robot control and motion for hardware acceleration. Second, we leverage a division deferring optimization in mass matrix inversion algorithm, which decouples reciprocal operations from the longest latency path to improve the performance. Finally, we present an inter-module DSP reuse methodology to improve DSP utilization and save DSP usage. Experiment results show that our work achieves up to 8x throughput improvement and 7.4x latency reduction over state-of-the-art RBD accelerators across various robot types, demonstrating its effectiveness and scalability for high-DOF robotic systems.
BibTeX
@misc{liu2025draco,
  author = {Xingyu Liu and Jiawei Liang and Yipu Zhang and Linfeng Du and Chaofang Ma and Hui Yu and Jiang Xu and Wei Zhang},
  title = {{DRACO: Co-design for DSP-Efficient Rigid Body Dynamics Accelerator}},
  year = {2025},
  eprint = {2511.08395},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.08395},
}
Figure