Hexagon's AEON humanoid moves sensor data over one fiber link using NVIDIA Holoscan

Hexagon Robotics has rebuilt the sensor pipeline in its AEON humanoid around NVIDIA Holoscan. Synchronized camera and depth data now reaches the robot's NVIDIA Jetson Thor compute module over a single fiber-optic connection. Hexagon reports initial measurements of less than 1 percent CPU usage per sensor stream and no GPU usage for camera processing.

The previous design wired sensors directly over MIPI to fixed ports on two compute units. All 12 ports were occupied, so sensor data had to be passed between the units. The cabling also created mechanical constraints where it ran through the robot's rotating neck.

The new architecture uses Holoscan as a common runtime, with Holoscan Sensor Bridge components handling inertial measurement unit connections and data transport. Hexagon built its own circuit board and FPGA software for AEON's head, which controls when each camera and depth sensor captures a frame so the observations line up. Data leaves the head through an NVIDIA-provided IP core and a fiber-optic cable. A custom receiver then passes it to the hardware-accelerated MGBE interface on Jetson Thor.

For automation engineers, the alignment matters because AEON is expected to detect and inspect objects, manipulate them, and capture the 3D environment for a digital twin. Those jobs depend on sensor frames that agree with one another in time.

Hexagon says the architecture supports peak network bandwidth of 100Gbit/s and eliminates buffer copies in the measured data path. Time-of-flight sensors need only a small amount of GPU computation. The company also says new sensors can connect through Thor's available network interfaces, directly or through a switch, without adding another central computing unit. Earlier, Hexagon processed images with GStreamer and NVIDIA DeepStream plugins, running AI models and classical computer-vision algorithms such as the feature detectors used in SLAM, while depth images moved between nodes through ROS 2.

Hexagon is now adding AI-based pre-processing of raw sensor data at each sensor and at the sensor-fusion level. The company says this improves object detection and categorization and gives AEON's motion planning and motion control modules more actionable data.