With NVIDIA Jetson embedded platforms, teams at the DARPA SubT Challenge detected objects with both high accuracy and high throughput.
Overview
The article discusses the performance of NVIDIA Jetson-based robots in the DARPA Subterranean Challenge, highlighting their capabilities in real-time object detection in challenging underground environments. It details the competition's structure, the technology used by participating teams, and specific achievements of notable teams like CERBERUS and Co-STAR.
What You'll Learn
1
How to leverage NVIDIA Jetson for real-time object detection in robotics
2
Why sensor fusion is critical in navigating unstructured environments
3
When to apply the YOLO model for object detection in robotics
Prerequisites & Requirements
- Understanding of robotics and real-time inference concepts
- Familiarity with NVIDIA Jetson platform and TensorRT(optional)
Key Questions Answered
What challenges do robots face in the DARPA Subterranean Challenge?
Robots in the DARPA Subterranean Challenge face challenges such as navigating unknown, unstructured, and uneven terrain, poor visibility, and lack of communication infrastructure. These conditions require advanced sensor fusion methods and robust robotic platforms to detect artifacts and objects of interest effectively.
How did Team CERBERUS achieve success in the competition?
Team CERBERUS successfully detected 23 out of 40 artifacts using four ANYmal robots and the Super Mega Bot. They utilized a modified YOLO model trained on 40,000 images, optimized with TensorRT, and deployed on NVIDIA Jetson AGX Xavier, achieving an inference rate of 20 Hz.
What technology did Team Co-STAR use for object detection?
Team Co-STAR used the medium variant of the YOLO v5 model for object detection, processing both RGB and thermal images. They trained their models using approximately 54,000 labeled RGB frames and 2,400 thermal images, achieving an inference rate of 28 Hz with their setup.
Key Statistics & Figures
Artifacts detected by Team CERBERUS
23 out of 40
This was achieved during the final competition using their robotic setup.
Inference rate of Team CERBERUS's model
20 Hz
This rate was achieved using the NVIDIA Jetson AGX Xavier.
Inference rate of Team Co-STAR's model
28 Hz
This was achieved using a combination of RGB and thermal imaging.
Technologies & Tools
Embedded Platform
Nvidia Jetson Agx Xavier
Used for real-time object detection and inference in robotic applications.
Machine Learning Model
Yolo
Modified for object detection tasks by teams in the competition.
Optimization Tool
Tensorrt
Used to optimize machine learning models for deployment on NVIDIA Jetson.
Key Actionable Insights
1Utilizing the NVIDIA Jetson platform can significantly enhance real-time object detection capabilities in robotics.By implementing NVIDIA Jetson, teams can achieve high throughput and accuracy in challenging environments, which is crucial for applications in search and rescue operations.
2Incorporating sensor fusion techniques is essential for navigating complex terrains effectively.Sensor fusion allows robots to integrate data from multiple sources, improving their ability to detect and interact with their surroundings, particularly in low-visibility scenarios.
3Optimizing machine learning models with TensorRT can lead to improved inference performance.Using TensorRT for model optimization ensures that deployed models can run efficiently on embedded platforms like NVIDIA Jetson, which is vital for real-time applications.
Common Pitfalls
1
Failing to account for the unique challenges of underground environments can lead to ineffective robotic designs.
Robots must be specifically designed to handle poor visibility and unstructured terrain, which requires careful planning and testing.
Related Concepts
Robotics In Search And Rescue Operations
Machine Learning Model Optimization
Sensor Fusion Techniques In Robotics