Here be dragons: physical AI, world models and robot brains
John, I'm only dancing. She turns me on. But I'm only dancing. Robot style.
I’ve been bullish for years about Nvidia’s building blocks for our robot and AI future, but at the same time, the faster one aspect progresses, the more likely we are to completely screw up our predictions of the rate of progress of the rest. One robot doing one thing in the real world is still highly complex and not well modeled. Our maps have far too many blank areas of unknown unknowns. Dragons. Maybe the answer is just across the ocean we’re paddling in. Maybe there’s an unknown continent in the way but we’ll find a way around it.
But many robots doing many things for many people is way beyond the scope of any believable maps we have. Rodney Brooks remains one of the strongest voices of realism in this discussion. Rod’s blog is compulsory reading for any roboticist, particularly his annual predictions scorecard.
Rocket Drew interviewed Nvidia’s VP for Robotics and Edge AI Deepu Talla (paywalled) who predicted that AI agents would bring about a ChatGPT moment for robots. Talla believes that there is a clear pathway for agentic AI to move from digital to physical. Robot companies currently collecting data will be able to utilize Cosmos simulation for training and testing which will lead to building general purpose robot brains.
When the general purpose robot brains exist, then agentic ai will act as the translation and traffic control layer, coordinating robot activity. The dangerous assumption dragon is that increased compute (and data) is all that is needed.
Aaron Edsinger, the CEO of Hello Robot, once asked Brooks for advice on building a robot company and was told that his dissertation and all the research was just one square in a sheet of graph paper and that creating a company that deployed robots in the real world required coloring in all the rest of the squares, with non research stuff.
Michael Harries, partner at the Robotics Hub and first check in Agility Robotics, talks about the ‘Risk Stack’ of robotics and deep tech. The compounding factors of complex deployments. We are more likely to run into dragons in the unknown unknowns of world models and physical AI than have a ChatGPT moment for robots.
Nvidia’s GTC has been crazy for robot after parties and we’ve co-hosted a couple of great robotics/physical AI events at Circuit Launch Mountain View, each attended by a few hundred people with interesting talks, panels and demos from Agibot, SoloTech students, hackathon winners and various Circuit Launch robotics companies.
Want to host an event or reach the robotics community?
Many companies are reaching out about hosting events, outreach or hiring and we have a great community network, of people looking for work, of investors interested in deal flow and of technologists interested in learning about the latest advances.
Would you be interested in having a company table at next Bots, Bevs & Devs or at the Robotics Career Fair on March 24? Or something customized specifically for your goals? Reach out here!
Business Insider got in touch with a Haidilao representative, who told us that the dancing feature is a “built-in, pre-programmed feature” of its robots. The restaurant has food delivery robots in action to help staff move dishes, but this specific model is geared toward entertaining customers and designed to “enhance the dining atmosphere.”
“Currently, this is the only Haidilao location in the US to feature this robot,” the spokesperson said, adding that the restaurant is “located in the heart of Silicon Valley, a perfect site to pilot this technology to align with the area’s innovative spirit.”
The call for proposals for Global ROSCon 2026 in Toronto is out.
More information here! Key dates you need to be aware of:
March 22, 2026 (UTC) Diversity scholarship applications are due! This is the only scholarship we offer, so plan accordingly
April 5, 2026 (UTC) Workshops proposals are due.
April 26, 2026 (UTC) Presentation proposals are due.
Boston Dynamics Spot robot dog for Sale
Good condition - robot is 4 years old, used in classrooms and business meetings. Includes tablet controller (UXV technologies), 2 batteries and charger.
Asking $35k obo
Mike Quan, text msg to: 617-500-7520
Silicon Valley Robotics Investor/Startup Database
Are you a startup fundraising atm? Fill out our “For Investors Eyes Only” sheet.
These investors write checks anywhere from angel and seed stage to $5-50 million.
Feedback from the startups - high quality inbound interest from investors.
Feedback from the investors - incredible depth and range of robotics startups.
If you’re an accredited investor and haven’t already signed up for the spreadsheet - Sign up here for great robotics deal flow!
Robotics News
Frontier Bio, one of our past neighbors here at Circuit Launch, just launched on Wefunder. Check out what they are doing with stem cells and bioprinting here!
Advanced Navigation has raised £83 million (Over €95 million) in Series C funding round to drive autonomy in high-stakes exploration in space and on Earth. The funding accelerates development of GPS-independent navigation sensors for extreme exploration and the company will establish PNT Centres of Excellence across Europe.
Set to debut in 2027, Advanced Navigation’s LUNA (Laser measurement Unit for Navigational Aid) will help Intuitive Machines land safely on the Moon, carrying NASA payloads.
This month, we launched LUNA’s terrestrial adaption - Chimera Land.
The sensor is supporting BHP to navigate Europe’s deepest underground mine.
The U.S Army validated it in a GPS spoofing and jamming battlefield.
Inside China’s robotics revolution - The Guardian
Unitree plans Shanghai IPO, testing interest in humanoid robots - Reuters
Horizon Robotics Invests In Physical AI company SynapX - Gasgoo
Mark Cuban says he thinks the humanoid robot push will fail in 5 to 10 years - Business Insider
Nvidia’s CEO says AI adoption will be gradual, but when it does hit, we may all end up making robot clothing - Fortune
NVIDIA and Global Robotics Leaders Take Physical AI to the Real World - NVIDIA
50 startups transforming industries with physical AI - Bessemer Ventures
A 24/7 solar farm-building robot just hit the market - electrek
AI Aims for Autonomous Wheelchair Navigation - IEEE Spectrum
Amazon acquires startup Rivr to test robots for ‘doorstep delivery’ - CNBC
Dental AI Robotics Startup Dentronic Raises $1M Pre-Seed from South Park Commons - WowTale
Building tomorrow: How Bedrock Robotics is changing the construction industry - The RobotReport
FieldAI and NVIDIA Omniverse: Building the Next Generation of Industrial AI - FieldAI
Uber to invest up to $1.25 billion in Rivian in robotaxi bet - SiliconValley.com
The Fight for Your Right to Repair - KQED
Send me your 2026 progresss, plans and posts for potential inclusion in Robots&Startups.
Robotics Events
Bay Area Material Handling & Robotics Expo, Career Fair and Reception at Cavalla Industries in SF
March 24 from 5pm to 9pm
RSVP on Eventbrite (or Luma) to attend the event
RSVP on Eventbrite to reserve a demo or recruitment table or short talk spot
Bots, Bevs & Devs at SVCTE in San Jose
April 1 from 6pm to 8.30pm
Details still tbc but see them first on the SVR Luma Calendar
SVR Leaders Circle in SF
April 6 - by invitation only
If you are CEO, CTO, Founder, Chief Scientist or Senior Researcher in a robotics company this is our peer networking event for robotics leaders. Save the date!
IEEE Hard Tech Venture Summit and Robotics Manufacturing Workshop at SRI International - April 16-17 - use “Csvrobo” for $90 discount
AMBayArea Summit 2026 (from the Association of Manufacturers Bay Area) - Chabot Space & Science Center - April 21
Keynote Speaker Andra Keay on Humanoids, Humans, and the Workforce
Revolution No One Is Talking About
More events:
Robotics Invest - Boston - May 26-27
ICRA 2026 - Vienna, Austria - Jun 01 - Jun 05
Automate
Deep Tech Week - SF - Jun 21 - Jun 25
Robot Fringe Festival - London - Jul 2
Robotics: Science and Systems - Sydney, Australia - Jul 13 - Jul 16
Open Sauce - San Francisco - Jul 17 - Jul 19
ROSCon Global - Toronto - Sep 22 - Sep 24
IROS 2026 - Pittsburgh - Sep 27 - Oct 1
IEEE Humanoids 2026 - Silicon Valley - Dec 6-9
The annual IEEE-RAS International Conference on Humanoid Robots is the internationally recognized prime event of the humanoid robotics community. The conference is sponsored by the IEEE Robotics and Automation Society and coordinated by the Technical Committee on Humanoids.
Join the organizing team at Silicon Valley Robotics - bots&beer@svrobo.org
Andra, the 'Risk Stack' framing from Michael Harries is exactly the right mental model here. Having deployed IoT and edge AI systems across industrial environments, I can confirm that the compounding complexity of real-world robotics is vastly underestimated by the 'ChatGPT moment' crowd.
The graph paper analogy from Brooks resonates deeply. In my experience with edge deployments, the non-research squares that need coloring include: connectivity reliability across heterogeneous wireless environments (a robot brain is useless if its telemetry link drops in a warehouse dead zone), device lifecycle management for fleets of hundreds of units each running different firmware versions, thermal management in non-climate-controlled industrial spaces, and regulatory compliance that varies by geography and vertical.
Deep Talla's vision of agentic AI as the coordination layer for physical robots is compelling in theory, but it assumes a connectivity and orchestration infrastructure that simply doesn't exist in most deployment environments today. The translation from 'digital agent books a flight' to 'physical agent coordinates with 50 other robots on a factory floor' requires ultra-reliable, low-latency mesh networking with graceful degradation -- not just better foundation models.
The real dragons on the map aren't in the AI models themselves. They're in the integration layer: getting sensor data from dozens of heterogeneous devices into a coherent world model, pushing OTA updates to robot fleets without bricking units mid-shift, and maintaining connectivity across environments where WiFi 6 barely penetrates the metal racking. These are deeply unglamorous infrastructure problems, but they're the ones that determine whether a robot demo becomes a robot deployment.
Rodney Brooks' realism is a gift to the industry. The companies that will win aren't the ones with the most impressive sim-to-real transfer -- they're the ones that have colored in all the boring squares around connectivity, fleet management, and operational reliability.
Andra, the Brooks graph paper metaphor is the best framing I've seen for why physical AI timelines consistently disappoint. I'd extend it further: it's not just that most squares are "non-research stuff" -- it's that many of those squares are invisible to the people drawing the roadmaps.
I deploy IoT and edge AI systems for a living, and the squares that consistently get left blank are connectivity and infrastructure integration. Everyone models the robot brain, the actuators, the sensors. Almost nobody models what happens when the WiFi drops in a warehouse corner and your AMR fleet loses coordination. Or when your edge compute node thermally throttles in a non-climate-controlled food processing plant. Or when you need to push a model update to 500 deployed robots across 12 sites with different network configurations.
Deep Talla's vision of agentic AI as the coordination layer for physical robots is technically elegant but assumes a communications substrate that largely doesn't exist in the environments where robots need to operate. Factory floors have RF interference from welding equipment. Agricultural fields have spotty cellular at best. Construction sites are dynamic RF environments that change daily as structures go up.
Michael Harries' Risk Stack concept captures this perfectly. The compounding risk isn't just technical -- it's that each layer of the stack (compute, connectivity, power, mechanical, environmental) can independently fail in ways that the other layers can't compensate for. A world model trained in Omniverse doesn't know that the facility's private LTE network drops packets under heavy load.
The companies that will navigate this are the ones treating deployment infrastructure as a first-class engineering problem, not an afterthought. Curious whether you're seeing startups that are tackling the connectivity/infrastructure layer specifically for physical AI deployments, or if it's still mostly left to the end users to figure out?