Scholarlink

We have developed a One Class Support Vector Machine (OCSVM) monitor to detect out of distribution (OOD) input to increase robustness of a state-of-the-art local path planning agent constituted of a Multi-Layer Perceptron (MLP) and trained through Deep Reinforcement Learning (DRL). To investigate deployment in an automated driving environment, we prototyped the system through High-Level Synthesis (HLS) on the ZCU104 System on Chip (SoC) Field Programmable Gate Array (FPGA). The proposed OCSVM FPGA design outperforms Jetson Nano CPU implementation in both latency (7 μs versus 941 μs) and energy consumption (10 μJ versus 474.62 μJ) per sample. A Raspberry Pi 5 device achieves 147 μs and 455.7 μJ. We showed that such an overhead for adding OOD monitoring in an automotive system is limited, since the MLP decision-making agent has a 1.5 ms latency and 1.1 mJ energy consumption per inference, on the same FPGA platform. The presented results–the first of this kind in literature–are promising and indicate directions for future research to further optimize the dedicated hardware (e.g., through quantization).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

FPGA Implementation of a Robust Decision-Making System for Low-Speed Maneuvering in Automated Driving

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Ammar Saad,
  • Hadi Ballout,
  • Francesco Bellotti,
  • Matteo Fresta,
  • Luca Lazzaroni,
  • Ossama Srour,
  • Riccardo Berta

摘要

We have developed a One Class Support Vector Machine (OCSVM) monitor to detect out of distribution (OOD) input to increase robustness of a state-of-the-art local path planning agent constituted of a Multi-Layer Perceptron (MLP) and trained through Deep Reinforcement Learning (DRL). To investigate deployment in an automated driving environment, we prototyped the system through High-Level Synthesis (HLS) on the ZCU104 System on Chip (SoC) Field Programmable Gate Array (FPGA). The proposed OCSVM FPGA design outperforms Jetson Nano CPU implementation in both latency (7 μs versus 941 μs) and energy consumption (10 μJ versus 474.62 μJ) per sample. A Raspberry Pi 5 device achieves 147 μs and 455.7 μJ. We showed that such an overhead for adding OOD monitoring in an automotive system is limited, since the MLP decision-making agent has a 1.5 ms latency and 1.1 mJ energy consumption per inference, on the same FPGA platform. The presented results–the first of this kind in literature–are promising and indicate directions for future research to further optimize the dedicated hardware (e.g., through quantization).