The Physical AI Dispatch, Issue #3
The Big Picture
Section titled “The Big Picture”If you build or deploy robots, this was the week that proved the infrastructure war is real. Embedded World 2026 in Nuremberg showcased an explosion of edge AI hardware purpose-built for robotics — from Qualcomm’s Dragonwing IQ10 shipping on real partner boards to Arm’s multi-robot coordination demos. Meanwhile, over $1 billion in fresh funding poured into robotics startups in a single week, MIT Technology Review declared physical AI “manufacturing’s next advantage,” and the EU quietly set a ticking clock on every robot fleet operator in Europe. With NVIDIA GTC kicking off Monday, the race to put intelligence inside physical machines is accelerating — and the companies that can deploy, manage, and scale that intelligence at the edge are the ones that will win.
This Week in Edge AI for Robotics
Section titled “This Week in Edge AI for Robotics”Embedded World 2026: The Edge Hardware Arms Race Heats Up
Embedded World 2026 (March 10—12, Nuremberg) was the biggest stage yet for edge AI in robotics. The dominant story: Qualcomm’s Dragonwing IQ10 platform is no longer vaporware — it’s shipping on real boards from real partners.
Advantech deepened its Qualcomm partnership by announcing the ASR-A703, a robotics and humanoid board built on the Dragonwing IQ10 series, targeting industrial AMRs and vision-language model workloads. SECO unveiled its SOM-COMe-CT6 module with up to 45 TOPS of on-device AI acceleration. ThunderSoft launched the TurboX IRB10 Development Kit for physical AI applications including humanoids and AMRs. GigaDevice showed edge AI solutions pairing GD32 MCUs with flash memory for humanoid motion control and sensor fusion.
Meanwhile, Arm demonstrated something arguably more interesting: Arm-powered robots, cameras, and industrial controllers autonomously discovering one another and forming coordinated, distributed AI systems. That’s not a demo — that’s a preview of how multi-robot systems will actually work at scale.
NVIDIA Jetson is still the 800-pound gorilla, but this week proved the Dragonwing ecosystem is real and shipping. For anyone deploying robots at scale, the days of a single-vendor edge strategy are numbered.
Physical AI Goes Mainstream: $1.1B+ Funding Week and MIT’s Seal of Approval
The capital markets and the media agree: physical AI has arrived.
Mind Robotics, a Rivian spin-out, closed a $500 million Series A co-led by Accel and Andreessen Horowitz to build an AI-enabled industrial robotics platform. Rhoda AI emerged from stealth with $450 million in Series A funding, training robots from hundreds of millions of videos. Sunday Robotics hit a $1.15 billion valuation with a $165M Series B for its household humanoid “Memo.” And Kraken Robotics closed a $402.5 million public offering to fund its $615M acquisition of subsea robotics group Covelya.
Simultaneously, MIT Technology Review published a major feature declaring physical AI “manufacturing’s next advantage,” citing a Deloitte survey showing 58% of business leaders already using physical AI and 80% planning to within two years. NTT DATA announced NVIDIA-powered “AI Factories” for enterprise deployment across manufacturing, healthcare, and automotive.
When $1B+ flows into robotics companies in a single week and MIT runs a physical AI cover story, the market is telling you something. The question is no longer “will robots deploy at scale?” but “who manages them when they do?”
GTC 2026 Preview: NVIDIA’s Physical AI Moment
NVIDIA GTC 2026 (March 16—19, San Jose) is the most robotics-heavy GTC ever. Jensen Huang’s keynote will cover the full stack — chips, software, models, and applications — with physical AI getting top billing. A dedicated panel features the CEOs of Skild AI, Waabi, and PhysicsX, three companies pushing the frontier of simulation-to-real transfer.
NVIDIA is launching OpenClaw, described as one of its fastest-growing open-source projects, with a build-a-claw event on the show floor. The recently released GR00T N1.6 model — integrating Cosmos Reason for step-by-step task planning — will be demonstrated on humanoid hardware. Behind the scenes, the open-source Newton Physics Engine (co-developed with Google DeepMind and Disney Research) is now available in Isaac Lab, giving researchers a free, high-fidelity simulator for robot training. 30,000+ attendees are expected across 700+ workshops.
NVIDIA’s simulation-to-deployment pipeline keeps getting deeper, but there’s still a conspicuous gap between “trained in Isaac Sim” and “running reliably on 500 robots.” That gap is called fleet operations.
EU Gets Serious: Robot Software Is Now a Safety Component
If you’re shipping robots to Europe (or plan to), pay attention: the new EU Machinery Regulation 2023/1230, effective January 20, 2027, formally classifies embedded software and AI as explicit safety components in machinery — including robots and cobots.
This isn’t just paperwork. The regulation requires continuous lifecycle compliance: qualification is no longer a one-time certification but must be demonstrated and maintained throughout the operational life of the machine. It works in conjunction with both the EU AI Act and the Cyber Resilience Act, creating a triple regulatory framework for AI-powered robots. Manufacturers have less than 10 months to prepare. For companies running robot fleets with OTA updates, model deployments, and edge firmware — compliance means audit trails, version tracking, and rollback capabilities across every robot in the field.
The EU is the first regulator to treat robot software as safety-critical infrastructure that must be managed continuously. This will drive massive demand for fleet-level compliance tooling.
Amazon Cuts 100+ Robotics Engineers — But the Robots Stay
Amazon laid off more than 100 white-collar positions in its robotics division in early March — the unit responsible for autonomous systems powering Amazon’s 1,200 global fulfillment centers. VP Scott Dresser called it “difficult but necessary.”
But this isn’t Amazon retreating from robotics. It’s Amazon rationalizing after shelving the Blue Jay humanoid project (killed quietly in January 2026 due to manufacturing costs and operational challenges) while doubling down on Orbital, its next-generation modular automation system. Amazon has deployed over 1 million warehouse robots and isn’t slowing down on automation — just redirecting engineering talent.
Blue Jay’s cancellation is a reality check for the humanoid hype cycle. Purpose-built automation that’s designed for specific environments is where real deployment scale happens — and that’s where fleet management matters most.
The Numbers
Section titled “The Numbers”Edge Hardware Watch
Section titled “Edge Hardware Watch”The 18-core CPU with up to 45 TOPS stole the show at Embedded World. Partners Advantech, SECO, ThunderSoft, and AAEON all shipped boards this week. Targets: humanoids, AMRs, and VLA inference. The most credible Jetson alternative yet. Source: Qualcomm
Blackwell-powered module with 4x energy efficiency gains. Jetson Thor remains the humanoid target with Figure AI and Agile Robots actively deploying. Expected to feature heavily at GTC this week. Source: NVIDIA
Showed at Embedded World as a low-cost solution for humanoid motor control and sensor fusion. Watch this space for sub-$50 edge compute in commodity robots. Source: Artiverse
AAEON’s GENE-R8K6 board at Embedded World targeting industrial vision and robotics. AMD positioning as the third option behind NVIDIA and Qualcomm. Source: AAEON
Open Source Corner
Section titled “Open Source Corner”Three major bridge projects emerged: ROBOTIS’s lerobot_robot_ros2_zenoh (LeRobot via Zenoh, no ROS install needed), lerobot-ros (lightweight interface for MoveIt2/ros2_control-based arms), and rosetta (full ROS 2 ⇔ LeRobot bridge). This is the convergence point where AI training frameworks meet production robot software. Source: ROBOTIS GitHub | ROS Discourse
Co-developed by NVIDIA, Google DeepMind, and Disney Research, Newton is now available in Isaac Lab. Free, open-source, high-fidelity physics for robot training. The simulation layer continues to commoditize. Source: NVIDIA Newsroom
Toronto, September 22—24. Diversity scholarships due March 22, workshop proposals due April 5, talks due April 26. Source: ROS Discourse
Zenoh will not become the default middleware for the ROS 2 Lyrical release; efforts restart to integrate the backend into Fast DDS. Important for anyone planning production ROS 2 deployments.
Open-source 6-DOF arm with ROS 2 Humble, LeRobot, and Isaac Sim support all targeting March 2026 availability. Sub-$500 hardware for embodied AI research. Source: VMME
One More Thing
Section titled “One More Thing”Google told CNBC this week that Intrinsic — the robotics software company it just absorbed from Alphabet — is “the Android of robotics.” Think about what Android actually did: it commoditized the hardware layer and moved all the value to the platform and services layer above it.
Now apply that to robots. If Google succeeds, the value won’t be in building the robot — it’ll be in the software that runs on it, the data that flows through it, and the fleet operations that keep it productive.
The companies building the robot equivalent of MDM, fleet analytics, and edge deployment tooling are sitting on the most valuable layer of the physical AI stack. They just don’t know it yet.