General Robotics says GRID onboards robots 360x faster and wrote 80% of its own code

General Robotics announced Wednesday that GRID, its robot intelligence platform, has begun engineering itself: onboarding robots, building simulations, creating skills, deploying them to hardware and debugging its own failures without a robotics engineer driving each step. 

The company calls it Auto Engineering.

It's essentially agentic AI, deployed to robot training, with pre-built harnesses to constrain and guide development. It automates much of robot training and physical AI development, making learning much, much faster.

According to General Robotics, that looks like skill transfer across form factors cut from three days to as little as 90 minutes. In other words, train on one robot type, works on many others.

The claimed numbers

(Percentages calculated on a calendar-time basis from the company's stated figures.)

What General Robotics demonstrated

The company published a technical post alongside the release, and it is more informative than the announcement.

The test case was a laboratory bench: two Flexiv Rizon arms needed to pick a test tube from a rack, pass it between them, and pour the contents into a beaker. 

GRID ingested the robot capabilities and controllers, built a hybrid simulation — a DFSPH fluid solver in NVIDIA Warp coupled to MuJoCo rigid-body dynamics, so the liquid behaved like liquid — composed a skill from existing perception and planning components, and deployed it.

Then it debugged itself. 

GRID found the assumed robot model disagreed with real controller measurements by 143 mm and 6.4°, and iterated the correction down to 5.7 mm and 0.68°. It caught motion commands going out at roughly 500 Hz instead of the intended 30 Hz and re-paced them to 29.7 Hz. When bad scene geometry started breaking motion planning, it benchmarked several depth models on the same camera capture and picked one that cut the tabletop's plane error from 28.0 mm to 1.2 mm.

That is calibration, control-loop and perception debugging ... the work that consumes deployment engineers for weeks if not months.

Then they demonstrated escalation by moving the beaker mid-pour. GRID trained a reactive policy on purely synthetic demonstrations. They asked for stirring with a thin rod and it decided it needed human demonstrations, requested them via GELLO teleoperation rigs, and fine-tuned a visuomotor policy. Constrained to a single phone video of a person swirling liquid in a flask, with no robot demos permitted, and it rebuilt the hand and flask motion in 3D and ran the interaction in simulation.

It was, essentially, training itself.

"One human video became a world of robot experience," the post says.

First task on a fresh robot: about four hours. Later skills on that same bench: as little as 10 to 15 minutes.

That's impressive.

For humanoid robots too?

The skill transfer General Robotics demonstrates is Flexiv Rizon → UR5e. Two industrial arms with similar morphology but different control interfaces. GRID chose a state-based representation with end-effector actions rather than joint-level control specifically so the policy would not be welded to one robot's kinematics.

That demo is not of a humanoid or a dexterous hand ... yet.

But the company states it is "applying this paradigm across various form factors and applications: industrial arms, dexterous hands, navigation robots, as well as humanoids."

That's interesting for anyone running a humanoid robot company and looking for faster ways to scale training.

The value for humanoids

The humanoid bet is that a general-purpose body unlocks general-purpose work. It depends on some level of general intelligence: robots that can do enough to do multiple tasks and to be quickly re-tasked as needed.

(Otherwise, traditional automation would be sufficient.)

General Robotics' bet is that the body is not the constraint but that deployment know-how and training time are the constraints.

"The capabilities are increasingly abundant," the technical post argues. "What remains scarce is the know-how to reliably deploy them."

If the intelligence layer genuinely becomes portable across hardware, then humanoid hardware differentiation compresses and training becomes much faster, much better, and much cheaper.

Hans Peter Brondmo, who ran Everyday Robots at Google X before it was shut down, is quoted in the release, and he has watched a very well-funded attempt at this fail:

"There's no single ChatGPT moment coming for robotics, just deployment-by-deployment progress, built on real-world data most companies aren't set up to capture. GRID is a platform that treats that problem as the product, which is exactly the unglamorous work that compounds."

Ecosystem partners named in the announcement include FANUC, for whom General Robotics is an Authorized System Integrator, and Galaxea Dynamics. Galaxea makes Kengo, a bipedal humanoid robot. Other current customers are described only by category: global top-5 companies across automotive manufacturing, port operations, energy generation, and food and beverage production, plus multiple government agencies.