There is a comforting story in artificial intelligence, and it goes like this: to make a machine smarter, you make the model bigger. More data, more parameters, more compute, and eventually the thing wakes up and does what you want. This has worked, roughly, for chatbots. General Robotics, a startup in Kirkland, Washington, is built on a suspicion that it will not work for robots - or at least not by itself - and that the suspicion is worth a company.
The company's argument is that a robot is not a text box. A text box has one job. A robot has a body, and a body has to see, decide, and move in a physical world that pushes back. General Robotics' position, stated on its own website with more confidence than hedging, is that "general intelligence emerges from rich composition of robot skills - not just larger models." The load-bearing word there is composition. You don't train one enormous brain to do everything. You build a lot of small, reliable skills - pick this up, open that, walk here - and then you snap them together like Lego.
What they actually sell
The product is called GRID, which the company describes as an "Intelligence Grid." It is a cloud-native platform that carries more than 40 pre-trained AI skills and pushes them onto robots - humanoids, robot arms, quadrupeds, wheeled bots, drones - through simple APIs. The pitch has three numbers in it, and the numbers are the whole point. Forty-plus skills. Deployment in under fifteen minutes. Twenty-five thousand concurrent robot requests from a single instance. And, crucially, no local GPU: the heavy thinking happens in the cloud, on-premise, or at the edge, but not necessarily on a graphics card bolted to the robot.
If that sounds less like a robot company and more like a software-infrastructure company, that is because it is one. General Robotics is not making the arms and legs. It is making the thing that tells the arms and legs what to do, and it would like to be that thing for robots it did not build, sold by companies it does not own. There is a partner integration with Trossen Robotics, whose hardware - the WidowX AI arm, bimanual mobile and stationary platforms - can run GRID skills without the customer wiring up GPUs and debugging drivers for a month.
"Real-world impact is constrained by the lack of a unified intelligence infrastructure."
The people
The founders are not tourists. CEO Ashish Kapoor spent seventeen years at Microsoft, latterly running autonomous-systems and robotics research, and helped create AirSim - the open-source simulator that a large fraction of the drone and self-driving research world used to test ideas without crashing real hardware. He co-founded the company with Dinesh Narayanan and Sai Vemprala. The broader team's resume reads like a tour of the last decade of applied machine learning: AirSim, Neural Simulation, VideoPoet, TensorFlow Object Detection, ClimaX. These are people who have shipped the unglamorous plumbing that other people's demos ran on.
A name that changed its mind
The company launched in 2023 under the name Scaled Foundations, backed by an undisclosed seed round from Khosla Ventures and E14 Fund. In 2025 it rebranded to General Robotics. This is mildly funny, because "Scaled Foundations" is a name that leans on the bigger-is-better thesis the company now politely disagrees with, and "General Robotics" leans on the composition thesis instead. Rebrands are usually cosmetic. This one reads like the company deciding, out loud, what it actually believes.
One of General Robotics' more interesting technical outputs is DreamControl, a workflow for teaching humanoid robots whole-body skills. It combines diffusion models - the same family of techniques behind AI image generators - with reinforcement learning, in a three-stage process, and hooks into GRID's hosted vision models for perception. The company says it wants to scale this to multiple humanoid form factors and thousands of skills, then compose those skills into more complex behavior. Which is the whole thesis again, restated as a roadmap.
Why anyone should care
The bet has a customer, and it is not a hobbyist. In April 2026, Accenture - a firm that does not invest in science projects - put money into General Robotics specifically to push physical AI into manufacturing and logistics. The stated problems are unromantic: workforce shortages, factory and warehouse productivity, rising operating costs. Accenture's Prasad Satyavolu framed robotics as an answer to "workforce constraints, challenged factory and warehouse productivity." The platform keeps customer data and intellectual property sovereign and trains robots in simulation using NVIDIA's Isaac Sim, which matters to enterprises that are nervous about handing their operations to a cloud.
None of this guarantees the thesis is right. Plenty of serious people are building the big-model version of robot intelligence, and they may win. But General Robotics has made a clean, falsifiable bet - that the bottleneck in robotics is no longer the hardware or the model but the software layer that makes any robot useful, quickly, at scale - and it has arranged an entire company around being the answer if that bet pays off.