What if the thing holding robotics back isn’t intelligence at all? The popular story says a single general foundation model will soon do for machines what ChatGPT did for text: one release, one viral moment, and suddenly robots are everywhere. Amulya Vishwanath thinks that story skips a few chapters.
Vishwanath spent seven years leading product marketing for NVIDIA’s robotics platforms before moving into venture capital. Her view, shared in a recent TechCrunch interview, is that robotics is in its “Lego era.” The basic building blocks, including actuators, robotic hands, software, and safety standards, are still being developed. You can’t build the castle while the bricks are still being designed.
As someone who works on agent architectures, I find that framing more useful than most of the hype. It moves the conversation from “when will the model arrive” to “what does the model need to stand on.”
Why the ChatGPT analogy breaks down
ChatGPT didn’t arrive in a vacuum. It landed on top of decades of standardized infrastructure. Text was already digital. Keyboards and screens were already in every home. Cloud compute was already a commodity. The model was the missing piece, and once it showed up, the rest of the stack was waiting to receive it.
Robotics has no equivalent substrate. There is no universal “keyboard” for a robot hand, no agreed-upon body plan, no shared standard for how a machine should behave safely around a person in a kitchen or a warehouse. A language model’s output is a string of tokens that any screen can display. A robot policy’s output is torque, force, and motion delivered through hardware that varies wildly from one vendor to the next.
That difference matters architecturally. In software agents, we can treat the environment as a stable API. In embodied agents, the environment includes the agent’s own body, and that body is still a moving target.
The four bricks still on the table
Vishwanath’s list maps neatly onto the layers any embodied agent depends on:
- Actuators. These set the ceiling on what any policy can do. A brilliant model driving an imprecise or fragile motor is still an imprecise, fragile robot.
- Robotic hands. Manipulation is where general-purpose robotics either works or doesn’t. Hands are mechanically complex, and without dependable ones, “general purpose” stays mostly theoretical.
- Software. The middleware between perception, planning, and control is still fragmented. Every team ends up rebuilding pieces that should be shared.
- Safety standards. This is the brick people forget. A chatbot that hallucinates produces a bad paragraph. A robot that misjudges a situation can hurt someone. Deployment at scale requires agreed rules, and those rules are still being written.
What strikes me is that only one of these four is primarily a machine learning problem. The rest are mechanical engineering, systems integration, and governance. That is a sobering reminder for those of us who assume progress flows mainly from bigger models.
Model-first versus component-first
There is a real tension here. One camp argues that a sufficiently general model will absorb hardware differences, learning to adapt to whatever body it is placed in. The other argues that without standardized, dependable components, there is nothing stable enough for a model to generalize across.
My read is that both are partly right, but the order matters. A foundation model for robotics is only as general as the variety of bodies and tasks it can be trained and tested on. If every robot is a one-off, every deployment becomes a custom integration project. Standard bricks are what turn a clever model into a platform. In that sense, the Lego metaphor isn’t a slight against the field. Lego became powerful precisely because the bricks were interchangeable.
Demos are not deployments
The visual evidence of progress is everywhere. At the Humanoids Summit in Tokyo on May 29, 2026, an exhibitor demonstrated a Unitree Robotics G1 humanoid for attendees. Events like this are valuable, and they show how far the hardware has come. But a polished demo on a show floor is a different test from months of unsupervised work in a messy, changing environment.
NVIDIA itself is pushing on many fronts at once. At CES in Las Vegas, Jensen Huang presented the company’s Rubin platform alongside work on open models and autonomous driving. Meanwhile, a Bloomberg Tech segment raised renewed concerns about circular financing tied to new NVIDIA deals, with discussion of AI agents and robots that will themselves be using computers. Ambition and capital are clearly not the bottleneck. The harder question is whether the physical layer can mature at the pace the investment implies.
What I’ll be watching
If Vishwanath is right, robotics’ breakthrough moment won’t look like a single model launch. It will look like quieter milestones: hands that hold up under repeated use, software layers that teams actually share, and safety frameworks that let robots leave the lab. Those signals are less exciting than a viral video, but they are the ones that tell you the bricks are ready.
From an agent-architecture standpoint, that is encouraging. It means the path forward is concrete. The intelligence is coming along. The body it lives in needs to catch up first.
đź•’ Published: