OpenAI intern connects Astra to a $200 robot arm to paint, signaling a coming DIY robotics boom

DIY robotics is closer than it looks

Thijs Simonian, a robotics intern at OpenAI, has been running experiments that connect OpenAI's models directly to a cheap consumer robot arm — and the results suggest the gap between "research demo" and "anyone can do this" is narrowing fast.

The arm behind him is a Hugging Face SO-100, fully open-source and 3D-printable. At roughly $200, it sits at a price point that makes hobbyist experimentation credible. Simonian connects it to Codex, gives the model a camera feed, and lets it paint. The first attempt was barely recognizable. By the fifth iteration, the progression was clear enough that he's planning to frame all five and call it "self improvement."

“I work here on the robotics team at OpenAI. I'm an intern and have been doing lots of exploratory projects... this arm that I'm using here, like, right behind me, this is a Hugging Face s o 100 robot, which is, like, open source, fully three d printable. You just need to get the actuators, which are, like, much cheaper. I think it at the moment... it's, like, around $200 or so, but, like, that's, like, incredibly cheap for robotic equipment and what you're able to do with it.”

How the control loop works

The naive approach — take an image, make one small movement, repeat — produces the best fine-grained results but is too slow to be practical right now. Simonian's workaround is a longer planning horizon: the model takes an image, writes a full plan in code, executes it, then monitors progress in the background every few seconds and makes small adjustments. Paintings took between one and two hours depending on method, with Codex running the inference.

Image resolution is a meaningful lever. He's been feeding the model 512p input rather than higher resolutions, and finds the model handles perception adequately at that quality. As control loops get faster and tighter, resolution will matter more, but for longer planning cycles it's a secondary concern.

The $200 robot and what it implies

Classical industrial robot arms run to thousands of dollars. At $200, the SO-100 changes the calculus for who can experiment. Simonian draws an analogy to the 3D printing wave — a period when consumer-grade hardware, combined with open-source software, pulled a whole community of builders into a space that had previously been gated by cost and expertise.

The same dynamic looks likely for robotics. Astra and Codex are already publicly accessible; cheap arms are arriving. Academic researchers are already starting to notice they can plug frontier models into physical hardware and get interesting results.

A humanoid robot built for under $2,000 — mentioned separately — reinforces the direction: the price floor for physical AI experimentation is collapsing faster than the capabilities are.

What's still missing

Speed is the binding constraint. The current loop is expensive and slow because each vision step costs a significant number of tokens. Faster inference, whether through smaller on-device models or dedicated chips, is what unlocks real-time applications — the difference between a robot that paints over two hours and one that could plausibly handle dynamic tasks like cooking or sorting mail.

Simonian's near-term interest is exactly those kinds of practical household tasks. The mail-sorting use case is representative: close the loop between a physical object in the world and a digital action, using a cheap arm and a vision model that can already identify and classify what it's looking at. No $50,000 humanoid required.

For anyone who wants to start now, Simonian points to the Hugging Face SO-100 as the entry point, notes there are open-source gripper modifications already circulating, and offers the obvious lesson from his own progression: the first painting looked nothing like the Golden Gate Bridge. Keep going.