BeetleBot: A Multi-Purpose AI-Driven Mobile Robot for Realistic Environments

Anh Nguyen, Erman Tjiputra, Quang D. TranJournal of Robotics & Autonomous Systems 2020 · DOI

TL;DR

This paper introduces BeetleBot, a versatile AI-driven mobile robot designed for realistic indoor, outdoor, and rough terrain environments. It addresses the limitations of specialized robots by offering a modular design and a robust AI ecosystem for autonomous navigation, perception, and human-robot interaction, all while providing extensive open-source simulation tools for rapid development.

What Problem Does This Paper Address?

The field of mobile robotics, while advancing rapidly, faces a challenge: many existing mobile robots are designed for very specific tasks, making them inflexible for new applications. Furthermore, integrating hardware, software, and AI for seamless deployment remains a significant hurdle. The authors aim to overcome these limitations by developing a new mobile platform, BeetleBot, that is adaptable to a wide range of scenarios and tasks, from autonomous delivery and exploration to human-robot interaction, and by providing a comprehensive simulation framework to streamline development and testing.

Methods

BeetleBot's design methodology is built on three core pillars:

  • Robust Mechanical Design: To achieve superior maneuverability, BeetleBot employs a rocker-bogie mechanism with six wheels, a departure from the common four-wheel designs. This mechanism allows the robot to traverse obstacles like rocks and debris by ensuring all six wheels maintain contact with the ground, enabling it to climb over them effectively. The design is also modular, allowing for easy customization of components like the wheel mechanism, cargo bay, and sensor suite.
  • AI-Driven Software Framework: BeetleBot boasts an "intensive AI ecosystem" designed to close the vision-control loop and enhance autonomy. This ecosystem is divided into three main applications:
  • Autonomous Navigation: Utilizes deep neural networks for navigation in both normal and complex environments.
  • Visual Recognition and Localization: Incorporates capabilities like face recognition, pedestrian detection, traffic light detection, and sensor fusion.
  • Human-Robot Interaction: Includes features for emotion estimation, reaction, and automatic question-answering. Crucially, this AI system is designed to be lightweight and achieve real-time performance, deployable on an NVIDIA Jetson Xavier embedded board.

  • Testing and Deployment: The authors emphasize the importance of rigorous testing. They have developed large-scale, open-source simulation models integrated with ROS (Robot Operating System). These simulations are designed to mimic realistic environments (indoor houses, outdoor cities, and natural caves) and are intended for rapid testing and deployment of robots.

Datasets

The paper does not explicitly detail specific datasets used for training the AI modules. However, it highlights the creation of large-scale, open-source simulation environments for testing and data collection. These environments include:

  • House environments: Approximately 130 objects in a 400m² area.
  • City environments: 275 objects spread across 3,000m².
  • Natural cave environments: 60 objects in approximately 4,000m².

These simulated environments are designed to be compatible with ROS and Gazebo, facilitating the collection of data for various learning-based methods, including autonomous navigation and object detection. The authors plan to release these environments and collected datasets.

Key Insights and Ideas

  • Rocker-Bogie for Versatility: Adapting the rocker-bogie mechanism, traditionally seen in planetary rovers, for a terrestrial, multi-purpose robot significantly enhances its ability to handle challenging terrains, a key differentiator from many indoor-focused mobile robots.
  • Modular Design for Adaptability: The modularity of BeetleBot is a critical design choice, allowing it to be easily reconfigured for different tasks without requiring a complete redesign, addressing the limitation of specialized robots.
  • Lightweight AI for Real-Time Performance: Developing an AI ecosystem that is both powerful (state-of-the-art accuracy) and lightweight enough to run in real-time on embedded hardware like the NVIDIA Jetson Xavier is crucial for practical autonomous systems.
  • Simulation as a Development Accelerator: The creation of extensive, ROS-integrated simulation environments is a significant contribution, offering a valuable resource for researchers to rapidly test and deploy their own AI algorithms and robot designs before physical implementation.

Key Results

The paper focuses more on the design and capabilities of BeetleBot rather than presenting extensive quantitative results. However, it states:

  • BeetleBot is capable of ultra-mobility, accurate localization, and autonomous navigation in normal, complex, and rough terrain environments.
  • The AI ecosystem achieves state-of-the-art accuracy while maintaining real-time performance, deployable on an NVIDIA Jetson Xavier.
  • The robot can traverse doorways, overcome obstacles, and navigate challenging terrains.
  • The paper references Fig. 2 and a video on their website for full demonstrations of BeetleBot's capabilities in various tasks and real-world scenarios, including sample trajectories and traffic detection.

Limitations and Future Work

  • Limitations: The paper mentions that due to page limitations, a full demonstration of the AI modules' performance is not included, directing readers to external resources. Specific quantitative performance metrics for the AI modules are not detailed within the paper itself.
  • Future Work: The authors express hope that their experience and open-source resources will benefit the robotic community. The release of their large-scale simulation environments and collected datasets is a key future contribution. The paper implicitly suggests that further real-world testing and validation of the AI modules in diverse scenarios will be part of ongoing work.

Summary updated

Abstract

We present BeetleBot, a new mobile robot that has been developed to operate in realistic environments. Differ- ent from previous state-of-the-art mobile robots, BeetleBot is designed based on the most recent advancements of actuator, mechanical design, and artificial intelligence. These advantages allow the robot to have powerful features including ultra- mobility, accurate localization, intensive AI ecosystem, and autonomous navigation in normal as well as complex and rough terrain environments. Based on its novel design, BeetleBot can be used in a variety of tasks including but not limited to autonomous delivery, exploration and surveillance, or human- robot interaction. Besides the development of the robot, we also build large-scale open-source simulation models that are fully integrated with ROS and can be used for rapid testing and deployment of any robots. Finally, we show the capability of BeetleBot and the performance of each individual AI module in various tasks and real-world scenarios.
BibTeX
@article{nguyen2020beetlebot,
  author = {Anh Nguyen and Erman Tjiputra and Quang D. Tran},
  title = {{BeetleBot: A Multi-Purpose AI-Driven Mobile Robot for Realistic Environments}},
  journal = {Journal of Robotics \& Autonomous Systems},
  year = {2020},
  doi = {10.31256/gn4vw9o},
}
Figure