Isaac ROS — Hardware-Accelerated VSLAM & Perception

What is Isaac ROS?

Isaac ROS is a collection of GPU-accelerated ROS 2 packages from NVIDIA. Instead of running perception algorithms on the CPU (slow), Isaac ROS uses CUDA to accelerate them on the GPU, giving 10-100x speedups.

Key packages:

Package	Function
isaac_ros_visual_slam	VSLAM — tracks robot position using camera
isaac_ros_nvblox	3D mapping of the environment in real-time
isaac_ros_object_detection	GPU-accelerated object detection (YOLO, etc.)
isaac_ros_depth_segmentation	Freespace detection for navigation
isaac_ros_image_proc	GPU-accelerated image processing pipeline

Visual SLAM (VSLAM)

SLAM (Simultaneous Localization and Mapping) answers: "Where am I, and what does my environment look like?"

Visual SLAM does this using a camera (instead of expensive LiDAR), making it perfect for humanoids with limited payload capacity.

# Install Isaac ROS VSLAM
sudo apt install ros-humble-isaac-ros-visual-slam

# Launch VSLAM with a RealSense camera
ros2 launch isaac_ros_visual_slam isaac_ros_visual_slam_realsense.launch.py

The output: a /visual_slam/tracking/odometry topic with the robot's estimated position and orientation.

VSLAM Pipeline

RealSense D435i
    │
    ▼
Left + Right Stereo Images ──────────────────────┐
    │                                            │
    ▼                                            ▼
Rectification (GPU)                    IMU Data (Accelerometer + Gyro)
    │                                            │
    └─────────────────────┬──────────────────────┘
                          │
                          ▼
            Visual-Inertial Odometry (GPU)
                          │
                          ▼
            /visual_slam/tracking/odometry
            (pose: x, y, z + quaternion)

Using VSLAM Output in Navigation

from nav_msgs.msg import Odometry

class PositionTracker(Node):
    def __init__(self):
        super().__init__('position_tracker')
        self.sub = self.create_subscription(
            Odometry,
            '/visual_slam/tracking/odometry',
            self.callback,
            10
        )

    def callback(self, msg: Odometry):
        x = msg.pose.pose.position.x
        y = msg.pose.pose.position.y
        z = msg.pose.pose.position.z
        self.get_logger().info(f'Robot position: ({x:.2f}, {y:.2f}, {z:.2f})')

nvblox — Real-Time 3D Mapping

nvblox builds a 3D voxel map of the environment from depth camera data, updated in real-time on the GPU:

ros2 launch nvblox_ros nvblox_ros_realsense.launch.py

This creates a 3D occupancy grid that Nav2 can use for path planning.

---

Next: Nav2 — Bipedal Navigation