Tag: Cognex

  • Qualcomm Dragonwing IoT Day: Six Ways Edge and Hybrid AI Are Working Today

    Qualcomm Dragonwing IoT Day: Six Ways Edge and Hybrid AI Are Working Today

    Demo stations, robots and a wide range of partners gathered to share the art of the possible at Qualcomm’s October Dragonwing IoT event in San Diego. AI has created renewed enthusiasm, innovation and investment within edge computing, IoT, and robotics. In many ways, the event was exactly what you would expect. Qualcomm laid out its strategy for industrial and edge AI, focusing on open development, purpose-built compute, and partnerships for robotics and machine vision. Qualcomm’s message was direct and consistent across the day’s presentations and meetings. It wants to expand its footprint in these markets using the Dragonwing portfolio and the engineering discipline it built in mobile and automotive.

    For enterprise buyers, the most useful insights came from discussions on how Qualcomm’s customers were using AI. The audience heard from Zebra, Cognex, Axon, GEOTAB, and Skydio, who didn’t speculate about what edge AI might do someday. These companies described what they are building now, why certain work has to happen on the device, and how it changes for their customers.

    Qualcomm’s bet: make the edge easier to build on

    Nakul Duggal, who leads Qualcomm’s industrial and embedded IoT and robotics efforts, described a company changing how it works. Qualcomm’s heritage is high-touch engagement with a relatively small number of large customers. The new model is developer-first. Developers can buy evaluation kits from distributors and use documentation, sample apps, and open software practices, including upstream Linux, Ubuntu, and Debian. Nakul shared that a customer should be able to order a kit and get productive without Qualcomm needing to pick up the phone.

    Qualcomm was also candid that the transformation is unfinished. One executive acknowledged the industry’s jokes that Qualcomm “may not be that easy to work with” and said that the company is working diligently to change this perception. In a question-and-answer session with media and analysts, Nakul Duggal said, “We’ve spent a lot of time on improving our prototyping-to-production process. We have traditionally always been a B2B company. Now, it’s how do we show up differently in the way that we work with the ecosystem (of developers).”

    Building Blocks of Qualcomm’s Dragonwing

    Qualcomm’s Dragonwing platform aims to offer modular, highly integrated building blocks that OEMs and developers use to create finished products faster and with less complexity. Instead of assembling separate processors, AI accelerators, connectivity chips, and real-time controllers, you get cohesive hardware and software foundations. They typically combine:

    • Application processors (Kryo or Oryon CPUs)
    • Graphics and On-device AI (Hexagon NPUs, ranging from ~1 TOPS on entry parts up to hundreds of TOPS on high-end robotics chips)
    • Real-time controllers (including RISC-V MCUs)
    • Connectivity (Wi-Fi, LTE/5G, Ethernet with Time-Sensitive Networking, etc.)
    • Vision/camera support, security, and extensive I/O
    • AI tooling that accepts pre-optimized models from Qualcomm AI Hub, models from Hugging Face, or a team’s own TensorFlow, PyTorch, or ONNX models. Teams can benchmark on cloud-hosted hardware before committing to a design. Edge Impulse serves teams that start with data rather than a model.
    • Arduino, acquired in October 2025, reported 280,000 Uno Q boards sold. Its new robotics-capable Ventuno Q offers a migration path to a production system-on-module built on Dragonwing IQ8.

    Hardware is paired with software stacks (Linux/Yocto, Android, Zephyr, SDKs, containerized microservices frameworks) and reference designs so customers can move from prototype to production more quickly while lowering bill-of-materials and board space. There’s also an ecosystem of more than 200 hardware partners and 35 distribution partners to carry designs from prototype to volume.

    Qualcomm CEO Cristiano Amon reiterated a common theme in tech today. AI is hybrid. The question is no longer edge versus cloud, but how much intelligence each device needs. “Once you accept that you may have on-premises or open-weight models, and you have big models in the cloud, by definition, you’re accepting that AI is hybrid,” he said, “and then how much AI you do on the edge is just the kind of computer you want to build.” He said, “It’s very easy for semiconductor companies right now to chase the shiny object of the data center.” He cited privacy, economics, and resilience as several reasons to run AI locally.

    Six things edge and hybrid AI are doing in production today

    1. Collapsing frontline workflows (Zebra Technologies)

    Zebra Technologies argues frontline workers have been underserved by technology. Its on-device models on Dragonwing combine barcode, image, video, RFID, voice, location, and other signals to understand what is happening around the worker. For example, it shared that proof of delivery for a global courier that delivers roughly 15 million packages a day. The AI captures only the data the operating procedure requires, excludes things like faces, and judges whether the proof image is good enough, so the driver no longer has to make that call. “You can collapse multi-step workflows into single instances of effort,” said James Poulton, the SVP and GM of Mobile Computing at Zebra Technologies. Zebra said this can add one or two deliveries per driver per day, which is a meaningful return at that scale. Zebra also discussed “token anxiety” as an overlooked driver of edge AI because the fear of runaway expenses when AI reaches 10,000 store associates is real. Zebra also described wearables and XR as its “next frontier,” which aligns with Qualcomm’s XR frontline-worker blueprint and its vertical priorities for retail, energy and utilities, advanced manufacturing, and heavy industry.

    2. Inspection at line speed (Cognex)

    Cognex CEO Matt Moschner described how manufacturers need real-time manufacturing defect detection at the edge. He said, “95% accuracy is not enough.”

    In 2022, Cognex developed a new inspection algorithm but had no processor that could run it at the required speed, price, thermal budget, and size. Moschner said, “Cognex had just invented a game-changing inspection algorithm. … What was the problem? We had nowhere to run it.” The gap Cognex described led to the Dragonwing-based In-Sight 3900.

    His example was seal inspection on coffee capsules at 30 milliseconds per part, about 2,000 inspections per minute. Cognex said its previous generation of computing could not reach that rate. Today, the company can reach a 99.9% defect detection at decision times measured in milliseconds.

    3. Situational awareness for public safety (Axon)

    Axon has built on Qualcomm since its Body 3 camera and now uses the platform across body, in-car, and fixed cameras for public safety. Two capabilities in the newly announced Axon Body 5 stand out. Officers can scan a driver’s license at the roadside and see warrant alerts without returning to the car. The Guardian capability streams audio from a recording camera to the cloud, where AI assesses the context of an encounter rather than listening for keywords like gun. It sends an alert when a situation appears to be escalating, so a person can decide whether to send help. This shows what hybrid AI looks like in public safety. The edge camera handles capture, connectivity, and location, while the cloud handles contextual analysis. Battery life remains a critical requirement because the camera must last a 12- to 16-hour shift. Blake Bullock, VP of Devices Product Management at Axon said, “It’s got to last the entire shift, or else we fail.” This vision of data capture and analysis has existed in the company for years, but we now have the underlying technologies to make it a reality. Bullock said, “We tried to do gunshot detection in 2018 on a body camera, and we failed. But we tried, we learned.”

    4. Trusted data from vehicles and assets (GEOTAB)

    Mike Branch, Vice President, Data & Analytics at GEOTAB, said the company connects about 6 million commercial vehicles and collects more than 100 billion data points a day. It uses this data to predict collision risk and breakdowns, check worksite safety (for example, whether workers wear hard hats), and track temperature-sensitive shipments. The theme was trust. Vehicle data used to be simple events such as a seat belt buckling. Now it includes computer vision inferences that must be correct, fast, and privacy compliant. “Trust is earned at the edge,” said Branch. GEOTAB has moved from an indirect relationship with Qualcomm to joint roadmaps, because its devices stay in vehicles for 10 years or more. It also makes its insights available inside Claude, ChatGPT, and Copilot via MCP, one example of how edge data will feed enterprise agentic AI. Qualcomm Dragonwing’s agentic skills and MCP tools act on host and target devices, so a developer can describe what to build instead of searching SDKs.

    5. Drones as autonomous infrastructure (Skydio)

    Skydio CEO Adam Bry described docked drones that launch autonomously or remotely, with no pilot on site. Bry said, “Drones have been on this journey from toys to tools to now being infrastructure.” In Oklahoma City, police launched a drone from a fire station rooftop dock after a train operator feared he had hit someone. The drone located the injured person in under a minute and guided responders to him. American Electric Power uses docked drones to inspect substations and found a burned-out pole component before it could cause an outage or fire. Caltrans uses them to inspect active landslides, so people don’t have to walk the terrain. These examples show how edge and AI can eliminate dirty, difficult, and dangerous jobs for humans.

    6. Robots that manipulate, not just move (PickNik Robotics)

    In September, Qualcomm announced its acquisition of PickNik Robotics, the company behind MoveIt, widely used open-source software for robot arms, and its commercial version, MoveIt Pro. PickNik co-founder Dave Coleman said MoveIt Pro now runs on Dragonwing IQ8 and Arduino Ventuno Q, with IQ10 support coming. Nakul Duggal said manipulation was the robotics capability least familiar to Qualcomm, which explains the deal.

    The deployments Coleman showed point beyond the factory. The presentation shared use cases such as automated car washes in Colorado where two KUKA arms scan each car’s shape, retail shelf stocking trained in simulation, airport baggage handling, cleaning in food plants and airport restrooms, and work with NASA and the U.S. Space Force. Coleman shared a concern that many enterprise buyers have expressed about the evolving world of physical AI: “How do you deterministically let the powerful AI models do their thing in the world … but not hurt anyone and not act irrationally?” Manipulation is less mature than the other examples, but it is deployed and will continue to improve over time.

    What it means for buyers in the world of physical AI

    My takeaway from the event was that the progress in edge AI, IoT and robotics is real and it’s accelerating. After all the hype of GPUs and data center buildouts, it’s clear to me that a wide range of chip technologies will be deployed to provide intelligence everywhere. Qualcomm and its partners demonstrated real examples of running AI on devices under tight constraints for power, latency, and cost. Hybrid AI was also a unifying theme where technology vendors can combine the power of the cloud with edge AI. Amon’s framing, Axon’s cloud-based Guardian analysis, and GEOTAB’s edge-plus-cloud model all point to hybrid rather than “edge only.”

    Thanks to Qualcomm for inviting me to be their guest at Dragonwing IoT Day. All event experiences for industry analysts and media were hosted, but no additional compensation was received. All opinions are my own.