OpenCV

After months of solo work, I'd like to present to you my biggest project yet, the conclusion of my bachelor studies in Robotics at Transport and Telecommunication Institute (TSI) — Isaac, a mecanum-wheeled mobile manipulator system capable of remote teleoperation, SLAM-based indoor mapping and goal-directed navigation. As a system, Isaac consists of multiple layers: ⚙️ Mechanical — an omnidirectional platform with four mecanum wheels, designed around a Lynxmotion A4WD1 Rover chassis and carrying a 4-DOF Waveshare RoArm-M2-S arm, translated into a digital twin for easy modification via 3D printing and for future DRL/VLA deployment in simulation. 🔌 Electrical — a custom 3S4P lithium-ion pack of twelve Samsung 18650 cells protected by a 40A BMS, feeding a single fused, switched 12.6 V bus that powers the Jetson, arm and drive motors for roughly 1.5–2 hours of active use in motion, with a dedicated charging port that reuses the existing arm power supply. 💻 Software — the ROS 2 software stack runs entirely on an NVIDIA Jetson Orin Nano, with closed-loop velocity control for the mecanum base, RTAB-Map for SLAM mapping via a wrist-mounted Intel RealSense L515 depth camera, and Nav2 for goal-directed navigation. On top of it all is a web interface accessible from any device to teleoperate Isaac, monitor errors and drop navigation goals straight onto a live map. I want to thank Transport and Telecommunication Institute (TSI) and my thesis supervisor Emmanuel Merchán for providing and funding every component of this project, giving me the experience of building a complex robotic system from scratch. It was a gruelling project with a few compromises made to meet the deadline — and still, I'm proud the world can finally see the fruits of my labour. I hope whoever picks this platform up next, to build their own ideas on top of it, shares the sentiment I have today. Demo video below — build photos per request. The project codebase can be found here: https://lnkd.in/dYaS4-HR I'm now actively looking for opportunities to bring this experience into real-world use. If your company is hiring, or you know someone who could use my robotics expertise — send me a DM or connect. Finally, my gratitude to the makers, communities and tools that made Isaac possible: NVIDIA Robotics · RealSense · Waveshare · Open Robotics · OpenCV · Anthropic. #Robotics #ROS2 #SLAM #Nav2 #MobileManipulator #Mechatronics #EmbeddedSystems #Jetson #Autonomy #RoboticsEngineering

I'm happy to share that I've earned the Certificate of Excellence in Image Processing and Computer Vision Algorithms from OpenCV. This certification has been an incredible learning experience, strengthening my understanding of image processing fundamentals, computer vision algorithms, and their real-world applications. It has helped me deepen my knowledge in a field that I'm truly passionate about. A special thanks to Satya Mallick for creating such high-quality content and making complex computer vision concepts easy to understand. #OpenCV #ComputerVision #ImageProcessing #ArtificialIntelligence #DeepLearning #MachineLearning #Python #AI #Learning #Certification

We won €2,000 in cash and €5,000 in Claude credits in the Intel Industrial Robotics Arm Challenge at the EUROPE EMBODIED 2026 Hackathon. 350+ applicants. 100 selected participants. 3 hardware tracks. 48 hours. 1 overall winner: 𝐑𝐨𝐁𝐨𝐨𝐬𝐭. We won our track with a standout live demo, pitched against the other track winners, and took 1st place overall. Thierry Popat, Pietro Dardano, and I decided to pivot from the given challenge and create our own. We wanted to solve a real industry problem, not simply make our robot complete a task. We realized that VLAs have an intrinsic limitation: they rely heavily on vision, but lack a physical understanding of the world around them. That is the problem we wanted to target by bringing perception of physical properties into VLM-driven robotics. Using the Franka Robotics Emika Panda robot, we enhanced one of the most common motions in industrial robotics: pick-and-place. We added weight-estimation-based sorting by using data from the robot’s joint torque sensors and feeding it into a custom-built neural network. This allows existing pick-and-place operations in factories to gain an additional sorting capability without requiring any new hardware, simply by using the sensors already present on the robot. In our demo, a 2D camera detected the bottles on the table and estimated their position. The robot then moved its end effector toward them and used its wrist camera with depth perception to refine the estimate before picking them up. Once lifted, the robot estimated the bottle’s weight and placed it in a predefined location based on whether it was light or heavy. The user interacted with the system through an API-based VLM, specifying where the objects should be sorted. The vision pipeline ran on OpenCV and YOLO models, and the only reason we could keep it efficient and low-latency was Intel OpenVINO. We exported our models to OpenVINO and ran inference on the powerful Intel Panther Lake machine, which handled all the vision before communicating with the main computer controlling the robot. This has been an amazing experience. I learned so much in such a short time, had the privilege of meeting so many talented people, and was extremely lucky to work with my two incredible teammates. A huge thank you to the sponsors who made this possible: Ignite Next, Intel, Anthropic, Amazon Web Services (AWS), ETH Robotics Club, and all the others. And a special thank you to ESRA - European Student Robotics Association, RoboTUM, START Munich and TUM International GmbH for delivering such an impressive event so smoothly. What I loved most was seeing some of the brightest minds in the European robotics ecosystem come together, united by the same passion for building. Can’t wait for the next one!!! #Europe #EmbodiedEurope #ESRA #intel #Anthropic #AWS #Robotics #Hackathon #Challenge #Building #Hardware #EmbodiedAI #PhysicalAI #Ingelligence #Munich #Hacking 

📷 OpenCV 5 just dropped, and it’s probably the most significant update to the library in years. If you work with computer vision or maintain ML pipelines, there are a few massive changes worth looking at. The biggest deal is the completely rewritten DNN inference engine. ONNX operator coverage jumped from around 22% in the 4.x versions to over 80% now. That means you can finally load most modern models straight out of the box without hitting missing operation errors. What I didn't expect: native LLM and VLM support. OpenCV 5 can now run models like Qwen 2.5, Gemma 3, and PaliGemma directly. It has its own built-in tokenizer and KV-cache, so you don't need any external dependencies to run a vision-language model. A few other practical highlights: • Native FP16 and BF16 support directly in cv::Mat. This is a huge quality-of-life fix that eliminates the manual type conversions we used to need when passing network outputs around. • A new Hardware Acceleration Layer (HAL v2.0) that brings 3x to 4x speedups on ARM for basic operations like resizing, plus native RISC-V support. • A total overhaul of the 3D vision modules, splitting point cloud processing and geometry into cleaner, distinct APIs. (For anyone maintaining older codebases, they also released 4.13 recently to keep the legacy track stable). Has anyone tried migrating an existing 4.x pipeline over to 5.0 yet? I'm curious to hear if the DNN and ONNX improvements are as smooth in practice as they look on paper. Docs: https://lnkd.in/eh6qK4tX Code: https://lnkd.in/eK7xNxnh OpenCV #OpenCV #ComputerVision #MachineLearning #EdgeAI

Interested, do join in. OpenCV

We are always excited to see what our partners create with Illumetry technology, and Roy Rodenhaeuser never stops surprising us with new ideas 💜 Recently showcased at the Pitch for the Future event in London, Roy presented an impressive exploration of intuitive interaction with volumetric datasets using Illumetry IO. The experience combines our holographic display with a tracked physical frame that acts as an anatomical slicing plane. As users move the frame through space, they can physically "slice" through volumetric medical data, creating a remarkably natural and immersive way to explore complex 3D scans. Sometimes, the simplest physical interactions create the most powerful experiences

We’re excited to introduce the All-New CodeCraft. CodeCraft is an AI-powered coding assistant designed specifically for hardware development. It helps makers, developers, educators, and students move from idea to working device faster. With CodeCraft, users can: • Generate code with natural language • Build directly in the browser • Compile firmware in the cloud • Flash hardware in one click Unlike general-purpose AI coding tools, CodeCraft is built around real hardware workflows — helping users reduce setup time and focus more on creation. 🎁 Launch Promotion Register now through July 13 to get 1 month of CodeCraft PRO for free. Explore CodeCraft: https://lnkd.in/gpCkeGg2 #CodeCraft #SeeedStudio #AI #HardwareDevelopment #IoT #EmbeddedSystems #STEMEducation #Maker

Computer Vision + Sports 🔥 🔥 🔥 Almost three weeks ago at Py4AI, I gave a talk about using open-source computer vision models in sports. Over 90 slides in 45 minutes. You should manage to watch the whole thing at 2x speed during lunch. ⮑ 🔗 youtube: https://lnkd.in/dfWuiaht #computervision #objectdetection #opensource #programming