Luxonis

Meet Luxonis Chat 2.0. Your AI co-pilot for everything OAK. https://chat.luxonis.com/ Need a code snippet to start a pipeline? Help debugging a specific error? Advice on which OAK device fits your project? Luxonis Chat pulls from our official docs and community knowledge to get you unstuck in seconds. Over the last two years, the assistant has handled nearly 40,000 public questions, and recent usage is running close to 2x the long-term average. So we rebuilt it. What's new in 2.0: - A polished, full-product UI with chat history and sidebar navigation - Quick-start prompt buttons for the most common workflows (Device Setup, - - Code & Examples, Help Me Choose OAK, and more) - Fully optimized mobile layout - Native dark and light mode - 40% faster response time on our internal benchmark - Smarter source handling: docs-first for accuracy, with forum context layered in when the docs don't fully cover an edge case We also recently released an agent-friendly markdown version of our full documentation at docs.luxonis.com/llms.txt. Drop that into ChatGPT, Claude, or any other agent to give it instant Luxonis context for custom coding workflows. This is just the first of many GenAI improvements coming to our platform. Try it out: https://chat.luxonis.com/

Got a custom-trained AI model? Here's how to deploy it to OAK 4 using Luxonis Hub. The last tutorial covered YOLO models specifically. This one is for any custom model you've trained, regardless of architecture. The walkthrough follows a solder defect detection model from upload to live inference on the device. One key recommendation: use the NN Archive format instead of raw ONNX. NN Archives bundle the model with its metadata and configuration, which makes conversion and deployment much cleaner. Resources: ☁️ Luxonis Hub: hub.luxonis.com 💾 OAK Examples: https://lnkd.in/enuZ78EJ 📄 NN Archive docs: https://lnkd.in/e4ij3qd5 🔬 Solder defect detection training blog: https://lnkd.in/eqMqdT4d

New tutorial: running YOLO neural networks onboard OAK 4 Deploying inference on the device itself cuts latency, removes the bandwidth tax of streaming video to a host or cloud, and simplifies system architecture. OAK 4 was built for this, with dedicated AI acceleration in the same package as the cameras. Bringing your own YOLO model is what makes the system actually useful in production, and OAK 4 makes it easy. Bring your model, convert it through Hub AI, and deploy it directly to the device. Watch the video walkthrough or follow along in writing: https://lnkd.in/eiTenB9y

Precision agriculture requires making split-second, plant-level decisions while moving at tractor speeds through dust and vibration. Farm-ING Smart Farm Equipment FlexCo is tackling this challenge head-on to transform sustainable farming. By developing high-precision solutions, like in-row mechanical weeders and spot sprayers, they are able to target individual plants instead of blanket-spraying entire fields. To make this practical at scale, they integrated Luxonis OAK-D S2 PoE cameras directly into their machines to handle the entire perception pipeline. Why they chose OAK: ⚡ Rapid Development: With a ready-made vision architecture and the DepthAI software stack, Farm-ING avoided building from scratch and was able to quickly test, iterate, and deploy new models. 🧠 On-Device AI: Running inference locally on the camera cuts latency, enabling the high-speed, high-resolution processing needed to make accurate decisions at real operating speeds. 🚜 Rugged Reliability: The IP-rated, single-cable PoE design is built specifically to handle the harsh environmental variables of outdoor farming. Smart enough for high-speed edge AI, rugged enough for the dirt, and ready to deploy out of the box. 📰 Read the full Farm-ING customer story: https://lnkd.in/eZbhcAHG 🛒 Shop the OAK-D S2 PoE: https://lnkd.in/eiY3q-VJ #AgTech #EdgeAI #ComputerVision #Robotics #Sustainability

We just wrapped our first international hackathon, and the results blew us away. The Luxonis team traveled to Milano to join Google Developer Group - PoliMi for the third edition of their GDG AI Hackathon. We equipped 12 teams (48 participants) with our new OAK 4 D setups and hosted the Best Vision Hack Challenge. For many of these students, it was their very first time picking up a Luxonis camera. What followed was a weekend of watching them build seriously impressive, highly sophisticated pipelines at record speed. Judging was tough. We evaluated every project across four axes: practical utility, creative device use, depth usage, and CV/AI model sophistication. The talent was so undeniable that we ended up calling out four teams instead of our planned three. Here are the winners of the Best Vision Hack Challenge: 🏆 1st Place: FORMA by Above and Beyond On-device YOLO pose estimation on the OAK 4 D for instant biomechanical coaching. 3D spatial deprojection replaces expensive coaches and wearables, delivering elite-level, privacy-compliant feedback directly to the athlete. Team: Raul Agolli, Aidana Akkaziyeva, Yvonne C., Federica Brasca 🔗 https://lnkd.in/ekXduNms 🥈 2nd Place: ParC by Los Bagigios 90 seconds in front of a camera gives a doctor a picture that helps recognize early signs of Parkinson's disease. One sensor, one session, catching what used to take multiple hospital visits over years. Team: Enrico WJ Yu, Lisa Wium, Carmine Pacilio, Luca Simei 🔗 https://lnkd.in/eMRKAxnw 🥉 3rd Place: SeeCure by Stunnatech A single OAK device replaces a fire system, people counter, earthquake detector, and evacuation tooling. Depth, IMU, people counting, and a fine-tuned YOLOv8n exit-sign detector, all running natively on camera. Team: Giuseppe Pisante, Edoardo Alberto Fumagalli Francesco Cerni Martina Raffaelli 🔗 https://lnkd.in/eip7nqBR 🏅 Honorable Mention: OAK-ULUS by 403 Forbidden A touchless surgical cockpit. Surgeons manipulate 3D CT/MRI scans without physical contact, preserving sterility. The system isolates the surgeon's gestures with millimeter precision, filtering out everyone else in the room. Team: Samuele Centanni, Francesco Della Casa, Lorenzo Di Maio, Matteo Bergamaschi) 🔗 https://lnkd.in/e6CpnGKn Huge shoutout to the event organizers for hosting such an incredible weekend, and a massive thank you to the Luxonis crew (Martin Peterlin, Matija Teršek, Jaka Škerl, and Klemen Škrlj) for making the trip, mentoring the students, and judging the final builds. #ComputerVision #EdgeAI #SpatialAI #Robotics #OAK4 #Hackathon #PoliMi

We just shipped major FPS improvements to LENS (Luxonis Edge Neural Stereo) in DepthAI v3.6.1. Classical stereo depth struggles with textureless surfaces, repetitive patterns, and difficult lighting. LENS solves this with learned feature matching and cost aggregation, running entirely on device with no external GPU. The numbers after optimization: - LARGE (768x480): 10 FPS to 22 FPS - MEDIUM (576x360): 26 FPS to 38 FPS - SMALL (480x300): 42 FPS to 56 FPS - NANO (384x240): 60 FPS to 85 FPS We also added four new high-resolution XL models scaling up to 1248x780 at 8.5 FPS for applications that need denser depth output. All of this runs on the OAK device itself. No cloud, no host GPU, no external compute. Full LENS Update: https://lnkd.in/esdS6PS7

AutoCalibration is now enabled by default in DepthAI 3.6.1. Stereo cameras drift over time. Vibration, thermal cycling, mechanical impact. When alignment shifts, depth accuracy degrades. Most OAK deployments never hit this problem, but long-running or mechanically demanding environments can. Starting with 3.6.1, DepthAI continuously monitors and corrects stereo calibration during runtime. No code changes, no setup, no calibration targets. Upgrade and it just works. Two modes: ON_START (validates calibration at boot) and CONTINUOUS (monitors and corrects throughout operation). Both can be controlled via a single environment variable if you need to tune behavior. This builds on the Dynamic Calibration library we shipped earlier, but the difference is you don't have to manage it yourself anymore. It's integrated into the runtime and on by default. Upgrade: pip install depthai==3.6.1 Full Release: https://lnkd.in/eGuQ5vyD

If you are building multi-view perception systems, motion capture setups, or high-speed tracking pipelines, timing is everything. Even a millisecond of lag between cameras can ruin 3D reconstruction and tracking accuracy. With the release of DepthAI 3.4.0, we’re bringing full FSYNC (frame synchronization) support to OAK 4 devices. Now, multiple OAK 4 cameras can operate on a shared timing signal, ensuring they capture frames at the exact same moment. This eliminates temporal offsets, making your multi-camera outputs directly comparable frame-by-frame. How it works: ⏱️ Precise Alignment: Eliminate motion blur and misaligned object positions across multiple streams. 🔗 Easy Daisy-Chaining: Link your OAK 4 devices together using our FSYNC Y-Adapters. ⚙️ Auto-Configuration: Hardware roles are determined automatically—the device without a cable on the IN port becomes the Master, and the rest act as Slaves. 💻 Simple Software Setup: Just assign the FPS to your Master camera, use the Sync node on your host to group frames, and start your pipeline. Ready to build tightly synchronized vision systems with minimal setup? Upgrade to DepthAI 3.4.0 today. 🗞️ Full technical deep dive: https://lnkd.in/d5nyyEwB 📄 FSYNC docs: https://lnkd.in/d7S-YVfQ 🛒 Shop the FSYNC Y-Adapter: https://lnkd.in/diFuC22k #ComputerVision #SpatialAI #Robotics #MotionCapture #EdgeAI #OAK4

Reflected light is one of the biggest challenges for spatial AI, especially when dealing with intense glare from sunlight or smooth, shiny surfaces like factory floors, glass, and metal parts. To solve this, we are introducing the OAK Filter Add-on Kit: quick, clip-on polarized sunglasses built specifically for the OAK-D-S2 PoE and OAK4-D cameras. Instead of guessing how to handle glare in your deployment, these kits allow you to instantly test how filtering light improves your stereo depth accuracy and texture detection. 👓 Test Instantly: Each kit comes with both Horizontal and Vertical polarizer attachments, so you can find the exact orientation that suppresses reflections in your specific environment. 🎯 Improve Depth Accuracy: Enhance the visibility of true surface features for much more reliable active stereo matching. 🔍 Better Inspection: Increase contrast in low-texture scenes and reveal surface defects (like scratches or dents) hidden under harsh glare. The Strategy: Validate Now, Integrate Later. We built these kits as the ultimate prototyping tool. Use the clip-on glasses to test and validate your pipeline today. Once you prove which polarization solves your environmental challenge, you don't need to buy hundreds of clip-ons. Contact our team, and we can natively integrate that exact filter directly into your camera lenses for mass production and fleet deployment. 🛒 Shop the Filter Add-on Kit: https://lnkd.in/dEtw8vCC 📰 Read the technical deep dive: https://lnkd.in/ddQ2dwqg #SpatialAI #ComputerVision #Robotics #EdgeAI #Manufacturing #OAK4 #AgTech

High Frame Rate Neural Networks on OAK4 at 480 FPS We’re introducing High Frame Rate (HFR) mode on OAK4, enabling up to 480 FPS — with neural networks running at the same speed. ⚡ What this means: • Real-time AI at ultra-high frame rates • Object detection (YOLOv6) at up to 480 FPS • Low-latency pipelines for robotics, automation, and more 📸 Current HFR modes: • 1920×1080 @ 240 FPS • 1280×720 @ 480 FPS This is an early preview on the RVC4 platform (IMX586), and we’re just getting started — more flexibility and features are on the way. 👉 Try the examples and see it in action: - Github: https://lnkd.in/dqpvCtTb - Blogpost: https://lnkd.in/dfwZrtbE High-speed vision is no longer just about capturing frames — it’s about understanding them in real time. #DepthAI #OAK4 #ComputerVision #EdgeAI #Robotics #AI #MachineLearning