Scholarlink

<p>Current state-of-the-art deterministic deep reinforcement learning (RL) algorithms that work on continuous action space may fail in problem-specific tasks such as autonomous driving. This work investigates why these algorithms perform poorly and proposes methods to remedy the identified problems. We present an end-to-end deep RL algorithm, navTD3, for autonomous driving tasks. navTD3 controls the throttle rather than the velocity to make our approach applicable to different autonomous vehicles. Moreover, classic deep-RL methods struggle, especially when training on autonomous driving tasks begins. In our algorithm, we develop a new exploration process to cover the state and action space of the task correctly and present a novel regularization technique called <i>Tanh</i> Regularization to prevent gradients from vanishing in the early stages of training. In a simulated office environment, we verify navTD3 on an autonomous ground vehicle equipped with an IMU sensor, three depth cameras, and a Bird’s Eye View map. We compare our approach with conventional deep RL algorithms, deep-deterministic policy gradients (DDPG), and twin delayed deep-deterministic policy gradients (TD3). navTD3 tackles the problems that DDPG, TD3, PPO and SAC suffer and outperforms them in terms of final performance, sample efficiency, and generalization ability.</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

navTD3: An End-to-End Learning Approach for Indoor Navigation

  • 期刊论文

  • 发表日期：

    2025年6月11日

查看全文
  • Furkan B. Mutlu,
  • Dogan C. Cicek,
  • Ramazan B. Guler,
  • Suleyman S. Kozat

摘要

Current state-of-the-art deterministic deep reinforcement learning (RL) algorithms that work on continuous action space may fail in problem-specific tasks such as autonomous driving. This work investigates why these algorithms perform poorly and proposes methods to remedy the identified problems. We present an end-to-end deep RL algorithm, navTD3, for autonomous driving tasks. navTD3 controls the throttle rather than the velocity to make our approach applicable to different autonomous vehicles. Moreover, classic deep-RL methods struggle, especially when training on autonomous driving tasks begins. In our algorithm, we develop a new exploration process to cover the state and action space of the task correctly and present a novel regularization technique called Tanh Regularization to prevent gradients from vanishing in the early stages of training. In a simulated office environment, we verify navTD3 on an autonomous ground vehicle equipped with an IMU sensor, three depth cameras, and a Bird’s Eye View map. We compare our approach with conventional deep RL algorithms, deep-deterministic policy gradients (DDPG), and twin delayed deep-deterministic policy gradients (TD3). navTD3 tackles the problems that DDPG, TD3, PPO and SAC suffer and outperforms them in terms of final performance, sample efficiency, and generalization ability.