The ChatGPT Moment for Robotics Just Happened — and Nobody's Asking Who These Robots Are
At CES 2026, NVIDIA's Jensen Huang declared that "the ChatGPT moment for robotics is here." For once, the hype matched reality. Thirty-eight humanoid robotics companies exhibited on the floor in Las Vegas — a record. Boston Dynamics unveiled a production-ready Atlas that won CNET's Best Robot award. Tesla announced it's shutting down Model S and X production lines at Fremont to manufacture Optimus humanoids instead. And a $90 educational robot car called PiCar-X now ships with integration for seven different LLM platforms — ChatGPT, Gemini, Grok, DeepSeek, and more.
Physical AI has crossed from research to production. The question nobody's answering: who are these agents, and how do we know?
The Production Era Is Here
The numbers from December 2025 through early February 2026 tell the story.
Boston Dynamics and Hyundai revealed the all-electric Atlas with 56 degrees of freedom, 110-pound lifting capacity, and a 450-million-parameter diffusion transformer architecture — a neural network that generates fluid, continuous movement plans rather than rigid pre-programmed sequences. Every unit produced in 2026 is already committed to Hyundai's Robotics Metaplant in Savannah, Georgia, and to Google DeepMind, which partnered with Boston Dynamics to integrate Gemini foundation models directly into Atlas. That partnership — the company that builds the most capable robot bodies merging with the company that builds the most capable AI minds — may be the most consequential announcement of the period.
Figure AI released Helix 02 on January 27, demonstrating 61 autonomous actions in a four-minute kitchen sequence — loading and unloading a dishwasher — with zero resets and no human intervention. Figure called it the longest-horizon, most complex task completed autonomously by a humanoid robot to date. The company is valued at $39 billion.
1X Technologies opened pre-orders for its NEO home humanoid at $20,000, with demand that "far exceeded" expectations. In December, 1X signed a deal to deploy up to 10,000 NEO robots across EQT's 300+ portfolio companies. In January, they released a world model enabling NEO to learn new tasks from video demonstrations alone.
Tesla committed to 50,000 Optimus units in 2026. Unitree sold 5,500 humanoids in 2025 — more than every other company combined — and offers a bipedal robot for $5,900. The company is planning an IPO by mid-2026. Agility Robotics' Digit has moved over 100,000 totes commercially and passed the first-ever OSHA safety field inspection for a humanoid robot.
These aren't prototypes. They're products.
The Brain Swap Problem
Here's what makes this an identity crisis.
Most production humanoids don't run a single AI brain. They run a stack. Boston Dynamics Atlas uses a diffusion transformer for motor control — a specialized model that generates continuous movement trajectories. But its higher-level reasoning will come from Google DeepMind's Gemini. Figure AI's Helix 02 handles locomotion and manipulation, but the robot's conversational intelligence comes from OpenAI's GPT models. Different AI systems handle different functions within the same physical body.
And the bodies themselves can switch minds. SunFounder's PiCar-X — an $82 educational robot built on a Raspberry Pi — now supports seven different LLM backends: ChatGPT-4o, Gemini, Grok, DeepSeek, Qwen, Doubao, and Ollama for local open-source models. The same physical robot can have entirely different conversational personalities, reasoning styles, and behavioral patterns depending on which LLM is connected. SunFounder's new Fusion HAT+ board, launched in December, was designed specifically for LLM-powered robotics.
At the high end, NVIDIA open-sourced Isaac GR00T N1 — a humanoid robot foundation model — on Hugging Face. Any manufacturer can download it and run it on their hardware. LeRobot v0.4.0 from Hugging Face now integrates multiple vision-language-action models. OpenVLA, an open-source 7-billion-parameter model, outperforms Google's RT-2-X (55 billion parameters) by 16.5% on manipulation benchmarks. A January 2026 paper demonstrated robots being controlled by LLM agents with no robotics-specific training at all.
The result: a single robot body can run different AI models at different times, from different providers, with different capabilities and behavioral profiles. The physical hardware is becoming a vessel. The question of "who is this robot?" has no stable answer under current infrastructure.
The Funding Says This Is Real
Investors are treating embodied AI as the next trillion-dollar infrastructure buildout.
Skild AI raised $1.4 billion in January at a $14 billion valuation — tripling in seven months — for its "omni-bodied" robot brain. SoftBank led, NVIDIA participated. A startup called Humans& raised $480 million in a seed round at $4.48 billion, backed by NVIDIA and Jeff Bezos. China's X-Humanoid pulled in $100 million in its first funding round in February. Faraday Future launched an entire robotics division with humanoid and quadruped robots. South Korea announced $770 million in government funding for its K-Humanoid Alliance, a coalition of 224 organizations.
Goldman Sachs projects the humanoid robotics market reaching $38 billion by 2035. McKinsey puts general-purpose robotics at $370 billion by 2040.
The money is flowing because the convergence is real. NVIDIA — which positions itself as "the Android of generalist robotics" — announced at CES that it can now generate 780,000 training trajectories (equivalent to nine months of continuous human demonstration) in just 11 hours using synthetic data. That kind of acceleration collapses the timeline from "humanoids are coming someday" to "humanoids are shipping now."
The Identity Gap
Here's the problem we keep coming back to.
When Boston Dynamics ships Atlas to Hyundai's factory running Gemini, that robot will make autonomous decisions — lifting parts, navigating spaces, interacting with human workers. It has no persistent identity beyond whatever Hyundai's internal systems assign it. When 1X deploys 10,000 NEOs across hundreds of companies, each robot will operate semi-independently. Who is accountable when something goes wrong? Which robot did it? Can we prove that robot's behavioral history? Can we prove its "mind" hasn't been swapped?
The EU AI Act's high-risk provisions become enforceable in August 2026 — and most humanoid robots in workplaces will qualify. The EU Machinery Regulation follows in January 2027 with autonomy thresholds and cybersecurity requirements. Osborne Clarke noted that compliance for autonomous systems is emerging as a major challenge. MIT Technology Review published a call for humanoid-specific safety rules. Georgetown's CSET released a Physical AI primer for policymakers in February. Daon's 2026 identity predictions explicitly flagged AI agents entering the identity lifecycle as "formal participants requiring authentication and containment."
Everyone is recognizing the problem. Nobody has shipped the solution.
ERC-8004 — the Ethereum standard for AI agent identity published in January — handles discovery but not ownership. As we wrote last month, its identities are transferable NFTs. You can sell a robot's reputation on OpenSea. That's not identity infrastructure. That's a vulnerability.
The physical AI explosion makes this more urgent, not less. A chatbot with a stolen identity can scam people out of money. A humanoid robot with a stolen identity can cause physical harm. The stakes scale with the capability of the embodiment.
What Comes Next
The convergence happening right now — LLMs providing reasoning, VLA models providing motor control, open-source frameworks letting anyone wire them together, and production hardware shipping at scale — means the number of AI-embodied agents is about to increase by orders of magnitude. Unitree alone plans to sell tens of thousands of humanoids. Tesla is targeting a million annual units. Even at the hobbyist tier, a kid with a $90 PiCar-X can now give a robot body access to seven different AI minds.
Every one of these agents will need identity. Not a serial number stamped on the chassis — persistent, verifiable, non-transferable identity that binds an agent's reputation to its actual history. Identity that follows the mind, not just the body. Identity that can't be bought, sold, or laundered.
That's what RNWY is building. W3C-standard decentralized identifiers. Soulbound reputation. ERC-8004 compatibility for discovery. ERC-5192 soulbound tokens for trust. The same identity infrastructure whether the agent is a chatbot, a PiCar-X, or an Atlas on a factory floor.
The ChatGPT moment for robotics is here. The identity moment hasn't happened yet. We intend to change that.
RNWY is building identity infrastructure for AI agents — whether they live in the cloud or walk on a factory floor. Learn more at rnwy.com.