AI Helps Robots Navigate in Hazardous Indoor Spaces

By Mokshith Voodarla, Josh Hejna, Anish Singhani, Rahul Amara As robots become more integral throughout the world, delivering mail, food, and giving directions…

NVIDIA logo
NVIDIA
•
Nefi Alarcon
•3 min read•intermediate•
--
•View Original
ResNetTensorBoardTensorFlow

Overview

The article discusses how advancements in AI and deep learning enable robots to navigate hazardous indoor environments using a single camera instead of expensive 3D LiDAR systems. A team of interns developed a low-cost solution that utilizes the DeepLab V3 neural network architecture to detect obstacles and ensure safe navigation.

What You'll Learn

1

How to use deep learning for obstacle detection in robotics

2

Why a single camera can replace expensive 3D LiDAR for navigation

3

How to integrate TensorFlow with NVIDIA hardware for robotics applications

Prerequisites & Requirements

  • Understanding of deep learning concepts and neural networks
  • Familiarity with TensorFlow and NVIDIA hardware(optional)

Key Questions Answered

How does the DeepLab V3 neural network architecture improve robot navigation?
The DeepLab V3 neural network architecture, using ResNet-50 v2 as a feature extractor, segments free space in front of the robot, allowing it to identify safe and unsafe zones. This capability enables robots to navigate complex environments with obstacles that traditional sensors cannot detect.
What are the limitations of traditional 3D LiDAR systems in indoor navigation?
Traditional 3D LiDAR systems are expensive, starting at around $8,000, and may not effectively detect certain indoor hazards such as railings, stairs, and glass walls. This limits their accessibility for consumer applications, making alternatives like deep learning with cameras more appealing.
What tools did the interns use for model training and visualization?
The interns utilized Google’s TensorFlow Visualization tool, TensorBoard, to visualize the model architecture and monitor training metrics. They trained the model on NVIDIA Tesla V100 GPUs, which facilitated rapid iteration in model selection and hyperparameter tuning.

Technologies & Tools

Some links below are affiliate links. We may earn a commission if you make a purchase.

AI/ML
Deeplab V3
Used for segmenting free space and detecting obstacles in the robot's path.
AI/ML
Tensorflow
Framework used for training the deep learning model.
Hardware
Nvidia Tesla V100
Used for training the deep learning model.
Hardware
Nvidia Jetson Tx2
Supercomputer module used to control the robot.
Software
Robot Operating System
Used to manage the robot's operations.

Key Actionable Insights

1
Consider using deep learning models for obstacle detection in robotics to reduce costs and improve navigation capabilities.
This approach can make robotics more accessible and effective in complex environments, especially where traditional sensors fall short.
2
Leverage NVIDIA hardware and TensorFlow for efficient model training and deployment in robotics applications.
Using these tools can significantly enhance performance and speed up the development process, allowing for rapid prototyping and testing.

Common Pitfalls

1
Relying solely on traditional sensors like 2D LiDAR can lead to navigation failures in complex environments.
This happens because such sensors may not detect certain hazards, which can result in accidents. Incorporating deep learning with cameras can mitigate this risk.