General Robotics CEO: we're wasting billions on physical AI chasing the wrong layer

Robot skill development is too slow

Robots do amazing things in demos. Getting those capabilities onto a robot that has to work a shift in a factory, a port, or a power plant is a very different story.

"That research-to-deployment gap is huge," General Robotics founder and CEO Ashish Kapoor told me on Humanoid (almost) Daily. "While we can see all these demos, if you are on the ground on the front lines, it's a pipe dream for you to go and implement those on robots."

Kapoor should know. 

He spent 17 years at Microsoft, where he led the Robotics and Autonomous Systems Group and helped create AirSim, the open-source simulator that became a go-to tool for generating training data for robots and drones. He started General Robotics in 2023, and last month the company made some very big claims about its GRID platform: 

1. robot onboarding cut from a month to as little as two hours
2. model ingestion from three days to 20 minutes
3. skill transfer across form factors from three days to an hour and a half
4. new skills created and deployed in as little as two days.
   

Those are company numbers, and they'll need to hold up across a lot more customers and a lot more robots. But how Kapoor says he got there is interesting. And his view on where the value in robotics will end up is even more so.

The core problem, he says, isn't the AI. It's the plumbing.

"I've been doing robots for many decades now, and we still build robots as if it were the 1990s or 1980s," Kapoor says. "That design hasn't changed."

In machine learning, frameworks like PyTorch and TensorFlow rapidly shrank the distance between a research paper and a working product. Robotics never got that moment. Instead, the stack fractured into silos: hardware, simulation, training data, fine-tuning, testing, validation, deployment. Each has its own companies, its own tools, and its own experts.

"Companies that do simulation, they only do simulation," Kapoor says. "Companies that do arms, they only do arms and will never, ever probably talk to a roboticist that has been working in the navigation space."

The result is that making a robot do something new takes a small army.

"The running joke in academia and researcher circles is that in order to make one robot move, you have a team of six graduate students behind it," he says. "And industries, they don't have the luxury of having those six top world-class experts or engineers."

AI agents to the rescue?

General Robotics' answer is what it calls Auto Engineering: essentially a team of virtual robotics engineers in the form of AI agents. Agentic coding has already changed how fast software gets built (I use it myself pretty much every day), but robotics is harder because you're jumping from bits to atoms. And agents can't do much in a stack that's scattered across dozens of incompatible pieces.

So for the past two and a half years, General Robotics did something most startups would be warned against. It built everything.

"We embarked on a journey to consolidate all the know-how that exists in robotics into a single monorepo," Kapoor says. That GRID monorepo includes robot form factors, simulators, the company's own models plus models built by others, deployment pipelines, inference code, and traditional planning code, all wired up with the hooks and APIs agents need to actually work on it.

(I suggested that this makes him Sauron forging the One Ring. He didn't object.)

Simulation is the clearest example of why that matters. Enterprises in manufacturing, logistics, and energy often spend months and millions building a simulation of the specific thing they care about, because no off-the-shelf simulator fits. With everything in one place, Kapoor says, agents can assemble one on demand.

"With a prompt, you could say, 'This is the scenario that I want to solve,' and agents go and say, 'All right, physics for this task is best if I use, say, MJX rendering, maybe Gaussian splats. Oh, by the way, we need some fluids as well. How about I go and create my own fluid solver,'" Kapoor says. "And within minutes, you have a simulator that never existed on the planet."

Real-world deployments with thousands of robots

The other driver is the shape of real-world deployments. Enterprises that want robots at scale want thousands of them, running 24/7 for months without breaking down.

"There is not a single manufacturer out there that can provide these numbers of units," Kapoor says. "So it is very clear that the future of robot deployment in enterprise is mixed mode. It won't be just one OEM."

That means mixed fleets: arms, old-school automation, quadrupeds for security, a few humanoids to fill gaps. Different makers, different form factors, and likely different AI models. Bringing a robot onto GRID means building a digital representation good enough to simulate, train, and test with, fine-tuning models for that body, and getting APIs and controls working so the whole stack can actually move it.

General Robotics works directly with OEMs to do that. 

They're partners, not customers. FANUC, Galaxea, and Flexiv are part of the ecosystem, and General Robotics is an authorized FANUC system integrator. Customers include government agencies plus large companies in automotive manufacturing, port operations, energy, and food and beverage. Investors include Nvidia, Khosla Ventures, Accenture Ventures, Construct Capital, E14 Fund, Shorooq Partners, and Valo Ventures.

Doing everything is a big bet. Kapoor knows it.

"The traditional wisdom in the startup world has always been: do one thing really well," he says. "I do not believe, for robotics, it holds … one of the fundamental reasons that we see so much failure in robotics is because it's easy to get sucked in and put a lot of capital in to develop capabilities, only to realize that the business surface area is not that much."

His hedge: "It's a big bet, but we do have support of agents now."

The bigger bet, though, is about where robot intelligence is headed. Kapoor sees three layers of training. 

1. At the bottom are the giant foundation models from the frontier labs. 
2. In the middle are robot foundation models: the VLAs and world models that have soaked up billions in venture capital. 
3. On top is post-training, which is making a specific robot useful for a specific job at a specific company.
   

"While billions of dollars have been pumped into the second layer, I have strong reasons to believe that the first and the second layer are going to merge," Kapoor says.

There's fresh evidence for that. 

OpenAI's GPT-6 Astra, released in early September, has been guiding robot arms in pick-and-place tasks with no task-specific training, and in some early tests has matched or beaten purpose-trained systems. Kapoor and some of his co-founders called this direction back in early 2023 while still at Microsoft, in a paper called ChatGPT for Robotics published just after ChatGPT itself arrived.

If he's right, that's a problem for anyone who has raised huge sums to build a robot brain from scratch. A company that spent billions on the middle layer could end up roughly matched by one that spent a few million on top of a frontier model.

"In our mind, always the magic was going to be on post-training, and it is ironic that so little capital has flowed there," Kapoor says. "And that's going to change in the next few months for sure."

Post-training is also where the customer is. And in most industries, whoever is closest to the customer captures the value.

"They get to call the shots on what OEMs succeed," Kapoor says.

If that's how it plays out, the robot makers may end up working for the integration layer, not the other way around.