Scholarlink

<p>This work presents a hybrid controller that combines a fast deep neural network (DNN) based controller with a high-fidelity model predictive controller (MPC) to facilitate lifelong learning capability. The hybrid controller is developed based on a novel combination of following three steps: (i) formulation of a stochastic optimal control problem (SOCP) using a simplified kinematic model of the vehicle, (ii) implementation of a DNN that learns to solve the SOCP from synthetic data, and (iii) integration of closed-loop online retuning via both direct learning and imitation learning. This strategy leverages the computational efficiency of the DNN based control and the reliability of the MPC, allowing the hybrid controller to continuously improve from expert demonstrations and live feedback. Experimental results in CARLA simulation demonstrate reliable path-following, with the DNN policy exhibiting progressive improvement over time. Theoretical analysis establishes the feasibility of the underlying SOCP and provides consistency guarantees for the DNN controller, while quantitative outcomes ensure lateral tracking errors below 0.3 m, yaw errors under 4.5 degrees, control efforts ranging from 0.2 to 2 units, and computation times around 50 ms across diverse urban environments in CARLA. Overall, the work highlights the potential of hybrid learning-based controllers for autonomous driving, offering enhanced adaptability and computational performance. </p>

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

A Lifelong Learnable Hybrid Controller for Autonomous Driving

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年12月17日

查看全文
  • Sandesh Athni Hiremath,
  • Naim Bajcinca

摘要

This work presents a hybrid controller that combines a fast deep neural network (DNN) based controller with a high-fidelity model predictive controller (MPC) to facilitate lifelong learning capability. The hybrid controller is developed based on a novel combination of following three steps: (i) formulation of a stochastic optimal control problem (SOCP) using a simplified kinematic model of the vehicle, (ii) implementation of a DNN that learns to solve the SOCP from synthetic data, and (iii) integration of closed-loop online retuning via both direct learning and imitation learning. This strategy leverages the computational efficiency of the DNN based control and the reliability of the MPC, allowing the hybrid controller to continuously improve from expert demonstrations and live feedback. Experimental results in CARLA simulation demonstrate reliable path-following, with the DNN policy exhibiting progressive improvement over time. Theoretical analysis establishes the feasibility of the underlying SOCP and provides consistency guarantees for the DNN controller, while quantitative outcomes ensure lateral tracking errors below 0.3 m, yaw errors under 4.5 degrees, control efforts ranging from 0.2 to 2 units, and computation times around 50 ms across diverse urban environments in CARLA. Overall, the work highlights the potential of hybrid learning-based controllers for autonomous driving, offering enhanced adaptability and computational performance.