FOURWEEKMBA TIMELINE · FROM THE NEWSROOM · UPDATED 16 NOVEMBER 2026
How to read the labels: OPENAI OpenAI’s own blog post or paper · PARTNER the other party to a deal, or the person involved, in their own words · REPORTED a named outlet’s reporting.
Era one: robots trained in simulation (2016 to 2020)
- 20 June 2016 OPENAI“OpenAI Technical Goals” lists “Goal 2: Build a household robot”: “We’re working to enable a physical robot (off-the-shelf; not manufactured by OpenAI) to perform basic housework.”
- 15–16 May 2017 OPENAIRoboschool, “open-source software for robot simulation”, then “Robots that Learn”: a system “trained entirely in simulation and deployed on a physical robot, which can learn a new task after seeing it done once”.
- May–July 2017 OPENAIThe Rubik’s Cube goal begins. OpenAI later wrote: “Since May 2017, we’ve been trying to train a human-like robotic hand to solve the Rubik’s Cube … We solved the Rubik’s Cube in simulation in July 2017.”
- 19 October 2017 OPENAI“Generalizing from Simulation”: dynamics randomization across “ninety-five properties” of the environment, and a new algorithm, Hindsight Experience Replay, after “a number of months unsuccessfully trying” standard methods on pick-and-place.
- 26 February 2018 OPENAI“Ingredients for Robotics Research”: eight simulated robotics environments, four for the Fetch arm and four for the ShadowHand, plus open code for Hindsight Experience Replay.
- 30 July 2018 OPENAI“Learning Dexterity” introduces Dactyl, a Shadow Dexterous Hand with 24 degrees of freedom “compared to 7 for a typical robot arm”. Training “used 6144 CPU cores and 8 GPUs … collecting about one hundred years of experience in 50 hours”.
- 15 October 2019 OPENAI“Solving Rubik’s Cube with a Robot Hand”, using a new method, Automatic Domain Randomization. The hand “solves the Rubik’s Cube 60% of the time (and only 20% of the time for a maximally difficult scramble)”.
- October 2020 REPORTEDOpenAI stops robotics research. A spokesperson told VentureBeat on 16 July 2021: “last October we decided not to pursue further robotics research and instead refocus the team on other projects.” Co-founder Wojciech Zaremba said on a Weights & Biases podcast, as quoted by VentureBeat, that a lack of data “was holding us back in terms of robotics”.
What Dactyl showed
The bet was that a robot could learn in a randomized simulator and then work in the real world without fine-tuning. OpenAI’s numbers show the cost of that robustness: learning to rotate an object without randomization took “about 3 years of simulated experience”, and in a fully randomized simulation “about 100 years”. Trained that way, Dactyl managed up to 50 rotations in a row on the physical hand; trained without randomization, 6.
Era two: partners, a new team, then models instead of a team (2024 to 2026)
- 29 February 2024 PARTNERFigure raises $675 million at a $2.6 billion valuation, with the OpenAI Startup Fund among its investors, and signs a collaboration agreement with OpenAI “to develop next generation AI models for humanoid robots”. OpenAI’s Peter Welinder, VP of Product and Partnerships, said in the release: “We’ve always planned to come back to robotics”.
- 4 November 2024 PARTNERCaitlin Kalinowski, who led Meta’s AR glasses hardware, posts: “I’m delighted to share that I’m joining OpenAI to lead robotics and consumer hardware.” OpenAI confirmed the hire to TechCrunch, which reported that OpenAI had recently started hiring research engineers for a robotics team.
- 4 February 2025 PARTNERFigure’s chief executive Brett Adcock on X: “Today, I made the decision to leave our Collaboration Agreement with OpenAI.” The collaboration lasted 341 days, by our count.
- 7 March 2026 PARTNERKalinowski posts “I resigned from OpenAI” and “I care deeply about the Robotics team and the work we built together”, citing OpenAI’s Pentagon agreement, as reported by Fortune.
- 29 October 2026 PARTNERSam Altman, on Stripe’s Cheeky Pint podcast: OpenAI has “a robotics program that we’re going to point at making lots of data centers”.
- 3 November 2026 OPENAIAn OpenAI post on how its hardware teams use its models says that work “spans robotics and compute infrastructure”.
- 13 November 2026 REPORTEDFortune reports, and OpenAI confirms, that the stand-alone robotics team has been disbanded. The research team moves under Chief Research Officer Mark Chen; the applied robotics and hardware teams move to the Applied organization under President Greg Brockman. A spokesperson said general-purpose models including GPT-6 Astra “have shown that can be directly applied to robotics”.
- 2027 (target) REPORTEDFortune says Altman is targeting a first public demonstration of OpenAI’s own humanoid robot in 2027. It is a target, not a scheduled event.
Why OpenAI says it changed course, each time
2020: not enough data
The spokesperson said “other approaches, such as reinforcement learning with human feedback, lead to faster progress”. Zaremba, as quoted by VentureBeat: “There [are] actually plenty of domains that are very, very rich with data. And ultimately that was holding us back in terms of robotics.”
2024: a path through partners
The Figure release framed it as combining “OpenAI’s research with Figure’s deep understanding of robotics hardware and software”, to help robots “process and reason from language”.
2026: general models do the work
OpenAI told Fortune its researchers had assumed at the start of the year that the route ran through video models such as Sora; since then, general models had proved directly applicable, so moving the researchers is the best way to accelerate progress towards general-purpose robots.
The evidence cited for the 2026 move is a simulation result: in an Nvidia study described by Fortune, an agent running on GPT-6 Astra scored 74.8% on a simulated manipulation benchmark against 51.0% for the best specialized robot models on that leaderboard. We have not read the study, and simulation scores may not carry over to physical robots.
- For robotics startups: OpenAI is now framing robot control as an application of one general model family, not a separate research bet. Fortune notes this challenges the case for robots with brains of their own.
- For OpenAI watchers: The 2026 change is a reorganization, not an exit: OpenAI says robotics remains core and Altman’s 2027 humanoid demonstration target still stands, per Fortune.
- For readers of the record: OpenAI’s own words cover the 2016–2019 work and the 2026 confirmation; the 2020 stop, the partnership’s end and the leadership moves come from statements quoted by outlets or by the other party.
What is not established
OpenAI has not published a post on either wind-down; both come to us through statements given to VentureBeat (2021) and Fortune (2026). Fortune dates the rebuild of a robotics team to 2024 without a month. No headcount has been reported for either team. Kalinowski’s and Adcock’s words are from their own posts as quoted by Fortune, TechCrunch and Decrypt; Altman’s are from the podcast as transcribed in our newsroom report. OpenAI’s own blog posts were read from archived copies because openai.com blocked automated access on 16 November 2026.
FAQ
Did OpenAI shut down its robotics team?
Twice, in different ways. It stopped robotics research in October 2020, a spokesperson told VentureBeat. In November 2026 it disbanded a rebuilt stand-alone robotics team and moved its researchers into the main research team and the Applied organization, Fortune reported and OpenAI confirmed.
What was Dactyl?
OpenAI’s system for a human-like robot hand (the Shadow Dexterous Hand), announced on 30 July 2018. It learned to reorient objects entirely in simulation and transferred to the real hand without fine-tuning; in 2019 a successor solved a Rubik’s Cube 60% of the time.
Why did OpenAI stop robotics in 2020?
Its spokesperson said other approaches, such as reinforcement learning with human feedback, led to faster progress; Zaremba said robotics lacked data compared with other domains.
What happened to OpenAI’s partnership with Figure?
It was announced on 29 February 2024 and ended on 4 February 2025, when Figure’s CEO said Figure would leave the collaboration agreement.
Is OpenAI building a humanoid robot?
Fortune reports that OpenAI says it is developing its own humanoid robot and that Altman is targeting a first public demonstration in 2027.
Related
Sources: OpenAI blog posts “OpenAI Technical Goals” (20 Jun 2016), “Roboschool” (15 May 2017), “Robots that Learn” (16 May 2017), “Generalizing from Simulation” (19 Oct 2017), “Ingredients for Robotics Research” (26 Feb 2018), “Learning Dexterity” (30 Jul 2018), “Solving Rubik’s Cube with a Robot Hand” (15 Oct 2019), read from Internet Archive copies of openai.com, and arXiv 1808.00177 and 1910.07113; Figure press release, 29 Feb 2024 (PR Newswire); VentureBeat, 16 Jul 2021; TechCrunch, 4 Nov 2024; Decrypt, 9 Feb 2025; Fortune, 7 Mar 2026 and 13 Nov 2026; Cheeky Pint podcast, 29 Oct 2026, as transcribed in our newsroom report. Quotes are verbatim from those sources. The 341-day count is our arithmetic.









