Perception Pipeline with SDG & Domain Randomization
End-to-end robotic perception pipeline with LiDAR SLAM, RRT exploration, and YOLOv8 object detection
Overview
Developed comprehensive end-to-end robotic perception pipeline integrating LiDAR-based SLAM, RRT exploration, and YOLOv8 object detection, achieving 95% validation accuracy with sub-50ms latency on NVIDIA Jetson Xavier NX through domain randomization and 5,000+ synthetic training images.
Technologies Used
Key Highlights
- Integrated LiDAR-based SLAM with RRT exploration algorithms
- Generated 5,000+ synthetic training images using domain randomization
- Achieved 95% validation accuracy with sub-50ms edge deployment latency
- Implemented unified ROS 2 architecture for distributed computing
Introduction
This comprehensive robotic perception pipeline project integrates cutting-edge technologies to create a robust autonomous navigation and object detection system. The pipeline combines simultaneous localization and mapping (SLAM), path planning, and real-time object detection in a unified ROS 2 architecture optimized for real-world deployment.
Skills Used
- Robotic Perception: LiDAR processing and sensor fusion techniques
- Computer Vision: YOLOv8 optimization and real-time object detection
- SLAM Systems: Mapping and localization algorithm implementation
- Edge Computing: TensorRT optimization for resource-constrained hardware
- Synthetic Data Generation: Domain randomization for robust model training
Project
The end-to-end perception pipeline addresses critical challenges in autonomous robotics by seamlessly integrating mapping, exploration, and object detection capabilities. The system utilizes LiDAR-based SLAM for real-time mapping and localization while RRT exploration algorithms enable autonomous environment discovery and navigation.
YOLOv8 object detection achieves 95% validation accuracy through comprehensive training on 5,000+ synthetic images generated using Isaac Sim’s domain randomization techniques. TensorRT optimization enables sub-50ms inference latency on NVIDIA Jetson Xavier NX, making the system suitable for real-time autonomous operations.
The unified ROS 2 architecture provides distributed computing capabilities with modular design principles, ensuring scalability across different robotic platforms. Software-in-the-Loop testing validates system performance across diverse operational scenarios, establishing a robust foundation for autonomous navigation applications in dynamic environments with reliable perception and decision-making capabilities.