Au-Zone | EdgeFirst

What is EdgeFirst AI?

EdgeFirst AI is Au-Zone Technologies' platform for deploying 3D spatial perception and 4D radar-vision fusion on edge devices. It enables autonomous machinery and robotics to perceive dynamic environments, supporting NVIDIA Jetson, NXP i.MX, and other embedded platforms.

What industries does EdgeFirst AI serve?

EdgeFirst serves construction equipment, precision agriculture, materials handling, autonomous mobile robots (AMRs), warehouse automation, mining, and marine cleanup systems where reliable spatial perception is critical for safety.

What hardware platforms does EdgeFirst support?

EdgeFirst supports NVIDIA Jetson (Nano Super, Orin), NXP i.MX processors (8M Plus, 93, 95), NXP Kinara Ara-2 NPU, web browsers, mobile devices, Au-Zone Maivin for vision, and Au-Zone Raivin for radar-vision fusion.

What is the difference between Au-Zone Maivin and Au-Zone Raivin?

Au-Zone Maivin is a vision-focused module for object detection and segmentation. Au-Zone Raivin combines 4D imaging radar with camera for robust 3D perception in challenging conditions like dust, fog, rain, and darkness.

What perception tasks does EdgeFirst support?

EdgeFirst supports object classification, 2D/3D detection, instance segmentation, semantic segmentation, panoptic segmentation, radar-vision fusion for 4D perception, and LiDAR-vision fusion for dense 3D point clouds.

What makes radar-vision fusion unique?

Radar provides reliable distance and velocity through dust, fog, rain, and darkness. Vision delivers fine-grained classification. The fusion engine combines both for industrial-grade perception when either sensor alone would fail.

How does EdgeFirst handle high-resolution and small object detection?

EdgeFirst ModelPack supports native 4K resolution (3840×2160) with NPU-optimized tiling inside the model, delivering better small object detection than traditional SAHI. Proven in The Ocean Cleanup's ADIS2 project.

How quickly can teams develop with EdgeFirst?

EdgeFirst delivers 20-50× faster development compared to building from scratch, with up to 9× reduction in annotation effort via auto-labeling. Production-grade models are trainable in weeks instead of months.