Nvidia-backed physical AI startup General Robotics says that it has built an AI-driven agentic platform for robot training that continually improves and teaches itself. The software, called GRID, cuts robot onboarding as much as 99.7%, reduces time to import new AI models by as much as 99.5%, and slashes skill transfer from one form factor to another – perhaps from a robotic arm to a humanoid robot – by as much as 97.9%, the company says.

In other words, General Robotics says it’s using AI to speed up robot skill training by orders of magnitude, helping robots learn new skills in minutes or hours rather than days, weeks, or months.

“People have been trying to deploy advanced robotic systems the same way for 50 years. With the rapid advances in AI, it’s time to rethink the approach,” Ashish Kapoor, founder and CEO of General Robotics, said in a statement. “Our GRID architecture is fundamentally different. It is the most efficient and flexible way to deploy and make robots useful by bringing decades of robotics knowledge, from research to software and AI models to simulations, data and hardware, into one agent-first platform with intelligence that compounds.”

In some cases, the company’s software trained a robot on a new skill by simply viewing one short video. The platform chose a video to simulation approach, the company says, reconstructed the motion in simulation, generated its own training data, then trained and deployed the new skill. In other cases, GRID found and fixed problems with existing AI models and hardware calibration, which it then corrected, among other problems.

General Robotics calls the approach Auto-Engineering.

The explicit model is coding agents. Software agents are getting really good, the company argues, because they were given harnesses — software repository access, execution feedback, the ability to revise — plus objective grading from tests, compilers and benchmarks. GRID has built four robotics harnesses that are meant to be the equivalent for physical AI. One ingests a given robot’s morphology and calibration, one assembles simulations for training, one creates skills and the final one deploys new skills and evaluates them on real hardware in the real world.

GRID learns and accelerates recursively, says General Robotics. A first task took some hours, but because each loop leaves artifacts behind — calibrations, trackers, control fixes and catalogued failures – later skills on that same platform deployed in as little as 10 to 15 minutes.

If this all works out in the real world – General Robotics says GRID is deployed with a top U.S. automotive manufacturer, the largest global port operator, a top oil and gas producer and multiple government agencies – this is big news.

Essentially, the company is saying that a major part of the cost, complexity and timeline in deploying robots is training, and that this new technology can significantly speed it up.

“The capabilities are increasingly abundant,” the company’s launch blog post says . “What remains scarce is the know-how to reliably deploy them.”

CEO Ashish Kapoor is not a first-timer making that claim. He ran Microsoft’s robotics and autonomous systems group and built AirSim, the simulator a large slice of the autonomy research world trained on. He co-founded the company as Scaled Foundations in 2023 with seed money from Khosla Ventures and E14, renamed it General Robotics in May 2025, and has since added Nvidia, Construct Capital, Shorooq and Valo Ventures. Accenture invested in April and pulled GRID into its manufacturing and logistics practice.

The company eats its own dog food: more than 80% of GRID’s own code is agentically written.