ROS2-Based Low-Cost Mobile Robot for Educational Assistance with Reactive Navigation and Semantic-Cached Language Processing

ROS2-based mobile assistant robot for educational environments, integrating offline NLP in Spanish, YOLOv8n object detection, and reactive navigation on Raspberry Pi 4B.

📄 Paper: Mobile Robot Assistant for Autonomous Navigation in Controlled Environments Using Natural Language Processing and ROS2 — Submitted to Robotics (MDPI), 2026.
👥 Authors: Sebastián Alexis Aucapiña, Nataly Cecilia Benalcázar, José Varela-Aldás and Ramiro Isa-Jara
🏛️ Institution: Universidad Tecnológica Indoamérica / ESPOCH — Ecuador

---

Overview

This project presents a low-cost educational mobile robot that combines autonomous reactive navigation with offline natural language processing in Spanish. The system runs entirely on a Raspberry Pi 4B using ROS2 Humble as middleware, making it accessible for resource-constrained educational environments such as public schools or universities in Latin America.

The robot operates in two modes:

• Education Mode — answers academic questions via voice interaction using a hybrid NLP architecture
• Navigation Mode — navigates autonomously using visual and ultrasonic perception

Total hardware cost: under $250 USD.

---

System Architecture

The system is organized into four functional layers:

┌─────────────────────────────────────────┐
│         Application Layer               │
│     (Mode Orchestrator - ROS2 Node)     │
├─────────────────────────────────────────┤
│         Processing Layer                │
│  Vision | Speech/NLP | Navigation       │
├─────────────────────────────────────────┤
│         Abstraction Layer               │
│     (ROS2 Interface Nodes)              │
├─────────────────────────────────────────┤
│         Physical Layer                  │
│  Raspberry Pi 4B | Sensors | Actuators  │
└─────────────────────────────────────────┘

ROS2 Nodes

Node	Function
usb_camera_node	Real-time image acquisition
object_detector_node	YOLOv8n-based object detection
object_follower_node	Motion command generation
ultrasonic_node	Obstacle distance measurement
imu_node	Orientation estimation (MPU6050)
motor_controller_node	PWM motor control (TB6612FNG)
voice_node	Speech recognition + TTS + NLP orchestration

---

Hardware

Component	Specification
Computing unit	Raspberry Pi 4B — 4GB RAM
OS	Ubuntu 22.04 Server (ARM64)
Motors	N20 DC motors — differential drive
Motor driver	TB6612FNG
IMU	MPU6050
Camera	USB Camera (Logitech C270)
Ultrasonic sensor	HC-SR04
Audio amplifier	PAM8403 + 8Ω speaker
Battery	7.4V LiPo 2200mAh
Voltage regulator	LM2596 DC-DC step-down

---

Software Stack

Component	Technology
Middleware	ROS2 Humble
Object detection	YOLOv8n (Ultralytics)
Speech recognition	VOSK (offline, Latin American Spanish)
TTS Online	Edge TTS
TTS Offline	Piper TTS
Local LLM	Qwen (quantized)
Cloud LLM	Trinity (online)
Semantic cache	Embedding-based similarity matching

---

Key Results

Metric	Value
Speech recognition accuracy (quiet)	98%
Speech recognition accuracy (noisy)	98%
YOLOv8n F1-score	0.975
Door detection recall	100%
Semantic cache accuracy	100%
Semantic cache avg. latency	3.8 s
Battery life — Education mode	96 min (1.46 A)
Battery life — Navigation mode	75.6 min (1.85 A)
Total hardware cost	< $250 USD

---

NLP Architecture

The NLP module uses a three-tier hybrid architecture to balance accuracy, latency, and energy consumption:

User Query (voice)
       │
       ▼
┌─────────────────┐     HIT      ┌──────────────────┐
│  Semantic Cache │─────────────▶│  Instant Response│
│  (embeddings)   │              └──────────────────┘
└────────┬────────┘
         │ MISS
         ▼
┌─────────────────┐
│  Local Model    │  Qwen (offline, quantized)
│  (Qwen)         │
└────────┬────────┘
         │ Complex query / no connectivity
         ▼
┌─────────────────┐
│  Cloud Model    │  Trinity API (online)
│  (Trinity)      │
└─────────────────┘

The semantic cache resolved 33.3% of queries without invoking any LLM, significantly reducing energy consumption and latency.

---

Installation

Prerequisites

• Raspberry Pi 4B with Ubuntu 22.04 Server (ARM64)
• ROS2 Humble installed
• Python 3.10+

Clone the repository

git clone https://github.com/SesSic/MobileRobotAssistantNLPROS2.git
cd MobileRobotAssistantNLPROS2

Install dependencies

pip install ultralytics vosk --break-system-packages
sudo apt install ros-humble-cv-bridge ros-humble-sensor-msgs

Build the ROS2 workspace

colcon build
source install/setup.bash

Launch

# Education mode
ros2 launch robot_assistant education.launch.py

# Navigation mode
ros2 launch robot_assistant navigation.launch.py

---

Dataset

Door detection model trained using transfer learning on the Door Dataset from Roboflow Universe (accessed March 2026).

---

Data Availability

The experimental data supporting the results of this study are available upon reasonable request: saucapina2@indoamerica.edu.ec / sebaaucaaa@gmail.com

---

Citation

If you use this work, please cite:

@article{aucapina2026mobile,
  title={Mobile Robot Assistant for Autonomous Navigation in Controlled Environments Using Natural Language Processing and ROS2},
  author={Aucapi{\~n}a, Sebasti{\'a}n Alexis and Benalc{\'a}zar, Nataly Cecilia and and Varela-Aldás José and Isa-Jara, Ramiro},
  journal={Robotics},
  publisher={MDPI},
  year={2026}
}

⚠️ DOI will be added upon acceptance.

---

License

This project is licensed under the Creative Commons Attribution 4.0 International (CC BY 4.0) license, consistent with MDPI open access policy.

---

Acknowledgments

The authors acknowledge the support of the Facultad de Ingenierías, Maestría en Robótica y Automatización Industrial, Universidad Tecnológica Indoamérica, Ambato, Ecuador.