Scholarlink

<p>Reliable terrain classification is a fundamental requirement for the autonomous navigation and safety of mobile robots operating in unstructured environments. Conventional approaches that rely on exteroceptive sensors, such as cameras or LiDAR, are often limited by environmental conditions, such as lighting and occlusion, whereas inertial-based methods provide only indirect estimates of surface properties. This paper addresses these limitations by presenting a novel instrumented airless wheel capable of direct terrain sensing. We introduce a manufacturing process that integrates piezoresistive carbon nanotube (CNT) sensors directly into the 3D-printed Thermoplastic Polyurethane (TPU) wheel structure, enabling the real-time capture of mechanical deformation signatures at the wheel-terrain interface. To process these signals, we propose a Recurrent Neural Network (RNN) architecture based on Bidirectional Gated Recurrent Units (Bi-GRU), which utilizes sequential deformation data and the robot’s linear velocity to identify terrain types. Experimental validation was conducted using a service robot traversing a 180-meter circuit composed of concrete, gravel, and grass surfaces. The proposed system achieved a classification accuracy of 97.67%, demonstrating that directly embedded sensing, combined with sequence-based deep learning modeling, offers a compact, high-precision solution for terrain-aware mobility without the need for complex external perception systems.</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

Terrain Classification Based on RNNs for Ground Robots Using a Custom Instrumented Wheel with Nano-Structured Sensors

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月16日

查看全文
  • Renato Silva Pereira Junior,
  • Luciano Quaresma,
  • Mário Delunardo,
  • André Cid,
  • Alan Rêgo Segundo,
  • Marcos Allan Reis,
  • Luiz Guilherme Barros,
  • Héctor Azpúrua,
  • Gustavo Pessin,
  • Gustavo Freitas

摘要

Reliable terrain classification is a fundamental requirement for the autonomous navigation and safety of mobile robots operating in unstructured environments. Conventional approaches that rely on exteroceptive sensors, such as cameras or LiDAR, are often limited by environmental conditions, such as lighting and occlusion, whereas inertial-based methods provide only indirect estimates of surface properties. This paper addresses these limitations by presenting a novel instrumented airless wheel capable of direct terrain sensing. We introduce a manufacturing process that integrates piezoresistive carbon nanotube (CNT) sensors directly into the 3D-printed Thermoplastic Polyurethane (TPU) wheel structure, enabling the real-time capture of mechanical deformation signatures at the wheel-terrain interface. To process these signals, we propose a Recurrent Neural Network (RNN) architecture based on Bidirectional Gated Recurrent Units (Bi-GRU), which utilizes sequential deformation data and the robot’s linear velocity to identify terrain types. Experimental validation was conducted using a service robot traversing a 180-meter circuit composed of concrete, gravel, and grass surfaces. The proposed system achieved a classification accuracy of 97.67%, demonstrating that directly embedded sensing, combined with sequence-based deep learning modeling, offers a compact, high-precision solution for terrain-aware mobility without the need for complex external perception systems.