On 1 June 2026 Sam Altman posted on X: AI should help humans in the physical world; in the short term OpenAI is building robots that assist skilled workers constructing future infrastructure, and in the long term it imagines a personal robot for everyone. The post announced OpenAI Robotics, closing a six-year loop from disbanding its robot team in 2020 to a loud return in 2026.
This is OpenAI’s renewed answer to the fundamental question of the path to AGI. Its robot work began earlier than most think. Between 2016 and 2019 it shipped the OpenAI Gym benchmark, the Roboschool simulation platform, and Dactyl, a dexterous hand that in 2019 solved a Rubik’s cube through reinforcement learning and automatic domain randomisation, proving simulation-to-reality training could work.
Then it went the opposite way. Around 2020 OpenAI disbanded the robotics team. Cofounder Wojciech Zaremba later said the root cause was the data bottleneck: physical-interaction data is scarce, expensive to collect and slow to iterate, while internet text and images are vast. Pouring resources into language models, the later ChatGPT, looked far smarter. The bet paid off, and OpenAI became the most valuable AI startup in the world.
The Figure AI breakup: why borrowing a brain failed
OpenAI did not fully exit. Through its startup fund it spread bets across Norway’s 1X Technologies, the US star Figure AI and Physical Intelligence. In February 2024 it joined Figure AI’s USD 675 million Series B and built a bespoke multimodal model; just 13 days later Figure 01 showed fluent natural-language interaction and object handling. Then in February 2025 founder Brett Adcock ended the partnership to build end-to-end robot AI in house.
Adcock was blunt: solving embodied intelligence at scale in the real world requires vertically integrating robot AI, and you cannot outsource the brain any more than you can outsource the hardware. The split taught OpenAI that investing in others was not enough. To truly understand the physical world it had to build the body and the brain itself.
From Worldsim to OpenAI Robotics
After the breakup, OpenAI quietly built a humanoid lab in San Francisco in early 2025 with about 100 data collectors training robot arms on household chores; by 2026 the lab had grown more than fourfold. Inside, a project called Worldsim, led by DALL-E and Sora creator Aditya Ramesh, aimed to let AI build a moving physical world in its head, computing how things fall, collide and get grasped.
On 1 June 2026 Altman declared OpenAI Robotics, grown out of Worldsim. The technical route is clear: simulation fidelity and sim-to-real transfer sit at the centre, so AI learns physical laws in a virtual world before that understanding moves into real robots. Altman set a roadmap: short term, robots that assist skilled workers in infrastructure, data centres, power grids and factories; long term, a personal robot for everyone.
The short-term target is itself a signal. OpenAI skipped home robots and warehouse bots and chose infrastructure building, the domain it knows best, where the AI compute boom has opened a massive labour gap in power and compute infrastructure that robots could help fill.
OpenAI’s embodied-intelligence roadmap
Phase one, 2016 to 2020, was exploration: from Dactyl to disbanding the team, proving sim-training-plus-real-transfer feasible before the data wall pushed it back. Phase two, 2020 to 2025, was investment: keeping a feel for the field via 1X and Figure AI, then learning that outsourcing the brain does not work. Phase three, 2025 to 2026, is return: from a humanoid lab to OpenAI Robotics, building body and brain in house.
OpenAI’s unusual angle is that it does not build the body first and then fit a brain. It lets AI understand the physical world first, then pours that ability into real robots. That reverses the path of Tesla and Figure, which start from hardware. If Worldsim truly lets AI learn physical laws in simulation and transfer cleanly, OpenAI could bypass the very data bottleneck that forced it to retreat six years ago.
In 2026, with stronger vision models, mature simulation tools and cheaper sensors, OpenAI returned with confidence, now holding the funding and talent to treat self-built robots as central to AGI. The challenge stays enormous: Figure’s breakup proved the gap between general models and robot control is real. Whether OpenAI can cross it will be answered the moment its first in-house robot leaves the lab.
More from the original report
Editor’s note: This is an adapted translation of the original OFweek Robot report. It has been trimmed and restructured for readability for an international business audience.
Translated and adapted from OFweek Robot (https://robot.ofweek.com/2026-09/ART-898890-8460-30702188.html).