Most active commenters
  • max_mclaughlin(3)

Show HN: Luxonis – OAK 4: spatial AI camera that runs Linux, with up to 52 TOPS

(www.luxonis.com)
16 points huntdunbar | 13 comments | 11 Dec 25 14:28 UTC | HN request time: 0.951s | source | bottom

Hey everyone. This is Hunter, CPO at Luxonis! We built OAK 4 (www.luxonis.com/oak4) to eliminate the need for cloud reliance or host computers in robotics & industrial automation. We brought Jetson Orin-level compute and Yocto Linux directly to our stereo cameras.

This allows you to run full CV pipelines (detection + depth + logic) entirely on-device, with no dependency on a host PC or cloud streaming. We also integrated it with Hub, our fleet management platform, to handle deployments, OTA updates, and collect "edge case" (Snaps) for model retraining.

For this generation, we shipped a Qualcomm QCS8550. This gives the device a CPU, GPU, AI accelerator, and native depth processing ISP. It achieves 52 TOPS of processing inside an IP67 housing to handle rough whether, shock, and vibration. At 25W peak, the device is designed to run reliably without active cooling.

Our ML team also released Neural Stereo Depth running our proprietary LENS(Luxonis Edge Neural Stereo) models directly on the device. Visit www.luxonis.com to learn more!

1. akouri ◴[11 Dec 25 14:57 UTC] No.46232152[source]▶
Congrats on the launch! What kind of models can you run on the device?
replies(2): >>46232980 #>>46234059 #
2. colbycook ◴[11 Dec 25 14:57 UTC] No.46232154[source]▶
I’m familiar with zed an Realsense but new to this, what are the differences here? How far does the on-board computer get me?
replies(1): >>46232637 #
3. filipproch ◴[11 Dec 25 15:02 UTC] No.46232235[source]▶
disclaimer I work for Luxonis/ Exciting to see this here after all the work that went into this, curious what people will create with it
4. CenekAlbl ◴[11 Dec 25 15:33 UTC] No.46232637[source]▶
Hi, there is too much more to OAK than Realsense or ZED so I will just try to tackle the Computer Vision/ML part. ZED and Realsense are basically 'simple' depth cameras, where ZED also requires external GPU to produce depth. With OAK 4 I can Use all sensor data (High-res images, Depth, IMU), run e.g. Object detection with YOLO, fuse it with depth to compute e.g. object distances and finally send those distance values out to my app/platform. You can use well established libraries like OpenCV, PyTorch and build your whole pipeline on-device so you save bandwitdth and do not need any host PC in most situations. OAK 4 is so powerful that you can run e.g. YOLOv8 Large at 85 FPS or DINOv3 at 40FPS. SImple models like Yolov6 and Text recognition or QR code decoding are running at 500+ FPS so there is plenty of room to combine them to get full vision stack.
5. aleshk ◴[11 Dec 25 15:57 UTC] No.46232980[source]▶
Thank you for your support! The device can run various AI and vision models; to get a grasp of what’s available, visit https://models.luxonis.com.
6. max_mclaughlin ◴[11 Dec 25 16:04 UTC] No.46233090[source]▶
Luxonis here - Incredibly proud of this launch. Use cases here are endless with the on-devie capability rivaling a Jetson Orin Nano. Excited to here how far and smart these devices can go.

Many more exciting launches coming soon - stay tuned for more on Neural Stereo

7. max_mclaughlin ◴[11 Dec 25 17:11 UTC] No.46234059[source]▶
You can run most standard CV models in parallel - including our own Neural Stereo depth estimation models (speed depends on the size you choose)

Here is a look at the performance of some standard models

YOLOv6 - nano: 830 FPS YOLOEv8 - large: 85 FPS DeepLabV3+: 340 FPS YOLOv8-large Pose Estimation: 170 FPS Depth Anything V2: 95 FPS

8. tylrtrmbl ◴[11 Dec 25 18:00 UTC] No.46234732[source]▶
What are some of the applications for this? I see a lot of agriculture and warehouse in the launch video but what problems does a stereo camera deliver value for?
replies(1): >>46235858 #
9. CenekAlbl ◴[11 Dec 25 19:18 UTC] No.46235858[source]▶
In agriculture the main ones are collision avoidance, human-machine safety - detecting and computing distance to an object nearby, then crop analysis - counting, measuring the dimensions and estimating ripeness. Warehouse adds use-cases such as inventory tracking, through volume estimation and barcode reading. Navigation (stereo VSLAM is more robust than mono), bin-picking, palletization, free-space estimation for cargo load. Those and many more require accurate depth estimation and strereo vision is the most universally applicable way of doing it, while being accurate and fast enough!
replies(1): >>46236519 #
10. max_mclaughlin ◴[11 Dec 25 20:14 UTC] No.46236519{3}[source]▶
Other verticals are smart cities, AMRs, manufacturing, security, and retail. Specific uses range from person detection, license plate recognition, demographic analysis, occupancy measurement, hazard zone monitoring, quality inspection, and street/road monitoring.
11. julienb0724 ◴[12 Dec 25 02:09 UTC] No.46240100[source]▶
Do users get root access? Can we build custom images?
replies(1): >>46243104 #
12. matijater ◴[12 Dec 25 11:28 UTC] No.46243104[source]▶
Not for the OS at the moment, but we offer this option to some enterprise clients.

Otherwise, you can easily connect via SSH and install any dependencies. Alternatively, you can use our ecosystem to build and run custom containerized applications.