Recursive self-improvement for physical AI

Published: October 2026
Author: Deepanshu Rohilla
Type: vision memo · 2 min read

Everything that has been digitised has been learned. Text, code, images, speech: anything that fits on a screen now has a model that gets better at it every few months.

But the digital world is a thin film over the physical one, and by our estimate it captures less than 0.001% of it. Under the film are the atoms. Every object that has to be lifted, turned, opened, stacked or carried is still moved by a hand.

If you can build something that improves itself in the physical world, you unlock realities humans can't yet imagine. It starts with a world where you don't have to lift a finger to move anything, and it ends with every atom around you being manipulated at will.

Every leap forward has come from learning beyond the limits of reality

1969

The limitApollo landing procedures had to account for uncertain computer alarms. One wrong abort call could end the mission.

The world built to learn inSimulated lunar landings that rehearsed alarms and abort decisions

What it unlockedApollo 11 flew through the real 1201 and 1202 alarms and landed on the Moon
2017

The limitGo AI learned from expert human games. Human knowledge capped where discovery could start.

The world built to learn inAlphaGo Zero played millions of games against itself

What it unlockedSuperhuman play with no human game data
2019

The limitRobot dexterity depended on real-world experiments. Physical data was slow and expensive.

The world built to learn inA robotic hand trained in simulations that grew more randomized as it improved

What it unlockedA real robot hand solved a Rubik's Cube
2026you are here

The limitRobots learn from demonstrations in the physical world. Expensive, slow and dangerous.

The world built to learn inInfinite worlds for robots to learn in

What it unlocks???
Next

A world where nobody has to lift a finger to move anything.
Then

Every atom around you, manipulated at will.

← Home Discovery-0 →

← Home

The Turing test for virtual worlds

Published: October 14, 2026
Authors: Argon Robotics team
Type: game 0.99 · 2 min

Hey, can you tell which ones are real?

Here are ten frames from the same robot stations. Some come from the real robot, and some from our simulation of it. Mark each one sim or real, then check your score. Every round draws a new set of ten from a pool of 98.

0 / 10 answered

Answer all 10 to see your score.

Real frames come from published evaluation runs of the ABC model. Sim frames are the same tasks rebuilt in our simulation and path-traced.

← Home Discovery-0 →

← Home

Discovery-0

Published: October 21, 2026
Authors: Argon Robotics team
Email: pranay@argonrobotics.ai
Video: dishwasher closing · recursive learning

Robot arms reaching for the open door of a commercial dishwasher — The robot closing a commercial dishwasher in a customer's kitchen, played at 2×. It had no real-world data for this task.

Robot arm pulling the dishwasher door down — The robot closing a commercial dishwasher in a customer's kitchen, played at 2×. It had no real-world data for this task.

Discovery-0 is our first demonstration that what a robot learns in our simulated worlds carries over to the real one. A model trained only on trajectories from simulation closed a dishwasher zero-shot in a live customer kitchen. We could not get this task working before, even with significant fine-tuning and teleoperated data collection.

Learning only inside the worlds

The robot then learned from trajectories it generated in those simulated worlds. For closing the dishwasher, we collected no real-world data at all: no demonstrations, no teleoperation, no fine-tuning on the customer's site.

Zero real-world data for this task.

Then we ran the model on the real robot in the real kitchen. It closed the dishwasher. The skill it had learned in our virtual kitchens transferred directly to the physical one.

The robot learned to drop the plate

The dishwasher was not the only thing the model figured out. In simulation, a plate got stuck at the edge of the rack, in a pose the gripper couldn't grasp well. Instead of retrying the same grasp, the model pushed the plate down, let it slide, and picked it up again from a better position.

In simulation, the model pushes the stuck plate down, lets it slide, and picks it up again. Overview camera on the left, wrist cameras on the right. Nobody showed it this recovery; it found it while exploring.

Dropping something on purpose so you can pick it up better is the kind of move people make without thinking, and the kind nobody writes into a dataset.

This discovery is not alone

There were more than 300 techniques that we traced back to sim explorations. This behaviour was not present in any data.

This loop is super useful. The virtual world proposes endless possibilities on model finding techniques to navigate the real world. Each discovery becomes something the model can build on, which is what we mean by recursive learning.

Technical breakdown

We will be releasing the complete pipeline, with the technical details of each block, in the coming days.

← HomeArgon Robotics · October 2026