NVIDIA-ISAAC-ROS/isaac_ros_nvblox

What it solves

Isaac ROS Nvblox provides a way for robots to reconstruct 3D scenes in real-time and create 2D costmaps for navigation. This allows robots to avoid obstacles using a vision-based approach, specifically integrating with the Nav2 stack for planning.

How it works

The system processes depth images, color images, and pose data (from sources like 3D LiDAR or depth cameras) using GPU acceleration. It leverages the nvblox C++ library to build a 3D reconstruction of the environment. This 3D data is then sliced into a 2D costmap for navigation.

To handle dynamic environments, it offers three modes of operation:

• Static: Assumes the environment is unchanging.
• People Reconstruction: Uses a DNN model (PeopleSemSegNet) to create segmentation masks for people, separating them into a distinct part of the reconstruction.
• Dynamic Reconstruction: Maps scenes containing general dynamic objects.

Who it’s for

Robotics developers building autonomous navigation systems that require real-time 3D mapping and obstacle avoidance using NVIDIA GPUs.

Highlights

• GPU Accelerated: Uses NVIDIA GPUs to compute 3D reconstructions and costmaps efficiently.
• Real-time Visualization: Streams mesh updates to RViz for live 3D scene monitoring.
• Dynamic Object Support: Specialized modes for mapping people and general dynamic objects.
• Nav2 Integration: Provides a costmap plugin for seamless integration with the ROS 2 navigation stack.