AI News

OpenAI Dissolves Robotics Team After 166 Days, Staff Folded into Core Model Line and Applications Department

HuxiuUpdated Nov 15 15:38

AI summary

According to Fortune, OpenAI has dissolved its independent Robotics team, which had existed for only 166 days; the researchers have been folded into the core model line, hardware staff have moved to the applications department, and robotics job postings have not yet been taken down. The article traces OpenAI's decade of reversals: in-house development in 2016, the halt in favor of language models in 2020, investments in 1X and Figure from 2023 to 2024, and the rebuilding of the team in 2025. NVIDIA's October tests showed that GPT-6 Astra achieved a 74.8% success rate on the RoboLab-120 benchmark, higher than the 51% of the specialized policy Lakeside, but it relies on Codex configuration, and direct action generation reached only 66.3%; over the same period, OpenAI's adjusted operating loss for the third quarter of 2026 reached $10.9 billion.

Why read this

AI score 63

The article reviews OpenAI's four adjustments to its robotics strategy over ten years, and presents general-model manipulation capability test results and loss figures as background for this consolidation.

Topics

Tags

Body · AI translation

[[**According to Fortune, OpenAI has dissolved its independent robotics team after 166 days, with staff merged into the main model line and the applications department; the move comes as general-purpose models demonstrate robotic manipulation capabilities, the company's losses widen, and it focuses on core business.**

**Key Points**

**1. Independent team dissolved after 166 days**

According to Fortune, the Robotics team was dissolved just half a year after its founding, with researchers merged into the main model line and hardware staff moved to the applications department; job postings have not been withdrawn.

**2. The robotics roadmap has been repeatedly adjusted over a decade**

In-house development began in 2016, pivoted away to language models in 2020; in 2023 it invested in 1X, in 2024 it partnered with Figure, in 2025 Figure terminated the partnership to develop in-house, and OpenAI subsequently rebuilt the team.

**3. Advances in general-purpose models weaken the need for a standalone team**

In October, Nvidia said GPT-6 Astra reached 74.8% on RoboLab-120, surpassing the specialized policy's 51%, but it relies on Codex configuration and achieves only 66.3% when generating directly; an internal video-model distillation step was skipped.

**4. Widening losses coexist with industry-wide escalation**

Adjusted operating loss for the third quarter of 2026 reached $10.9 billion; the company is preparing for a possible IPO and focusing on its core. Meanwhile, NEURA, Gemini Robotics 2, DeepSeek, and Figure continue to invest; practitioners say GPT cannot replace embodied models, and OpenAI plans to showcase its own robot in 2027. Original 2026-11-15 14:50 OpenAI dissolves robotics team Song Sihang Quick Take

166 days ago, Sam Altman announced the founding of OpenAI Robotics and released 19 job postings at once, which was once called by outsiders the "ChatGPT moment for humanoid robots." And just recently, OpenAI dissolved this newly established robotics team.

According to Fortune, this Robotics team, founded only half a year ago, has been dissolved, with researchers merged into the main model line and hardware staff merged into the applications department. At this point, the robotics positions on OpenAI's careers page, with annual salaries of up to $485,000, have still not been taken down.

During these 166 days, investment across the embodied AI industry has continued to escalate. NEURA announced up to $1.4 billion in financing; Google DeepMind released Gemini Robotics 2 to advance full-body robot control; DeepSeek spent about 141 million yuan to participate in Unitree Robotics' IPO strategic placement; and Figure reached a partnership with Nscale, planning to use up to 100,000 GPUs to train robot models. Nearly every company is preparing for the next round of expansion.

Against this backdrop, OpenAI's decision sounds dramatic, but it is also very much in line with OpenAI's consistent style. First, this is not the first time OpenAI has dissolved a star team; the last time was Sora.

That team, once regarded as OpenAI's multimodal ace, was scaled back and merged into the main line under Altman's directive to "cut all side quests." OpenAI has an almost cold habit: no matter what happens in the outside world, and no matter how promising the direction looks at the time or how strong the team's paper capabilities are, as long as it is internally judged no longer the shortest path to AGI, or as long as a general-purpose model can soon handle that direction as a byproduct, the independent side quest will be merged into the main quest, and the independent team will lose its reason to exist.

Dissolving the robotics team follows the same logic. And in fact, these 166 days were not an impulsive, messy retreat. Looking back 11 years, this is not the first time OpenAI has cut an independent robotics team.

What happened over these ten years?

OpenAI's first foray into robotics came six years earlier than ChatGPT.

In 2016, the newly founded OpenAI listed several things it wanted to do in a research plan. One of them was to build a robot capable of completing basic household chores. At the time, the company had not yet focused its main efforts on language models. Robotics, games, and reinforcement learning were all directions in which it explored artificial general intelligence.

That year, OpenAI released the reinforcement learning toolkit Gym; the following year, it launched the robot simulation tool Roboschool. The team hoped to first let machines learn movements in simulated environments, then apply the learned capabilities to real hardware. Compared with having robots trial-and-error repeatedly in a lab, simulation can run large numbers of experiments simultaneously and does not risk breaking equipment.

In 2018, this approach produced a striking result called Dactyl. It was a dexterous robotic hand whose task was to adjust the orientation of a block in its hand. A block turning in a human hand looks effortless, but it is very difficult for a machine: how the fingers coordinate, where force is needed, and how to recover after the block slips all have to be solved. OpenAI trained the control policy in simulation, then transferred it to a physical robotic hand.

A year later, the task changed from turning a block to solving a Rubik's cube.

On October 15, 2019, OpenAI released a video of Dactyl solving a 3x3 Rubik's cube. To adapt the policy trained in simulation to reality, the team continuously varied the physical parameters in the simulated environment, including friction, object mass, and the motion conditions of the robotic hand. This approach, called "automatic domain randomization," was meant to teach the system to cope with changes that could not be anticipated one by one during training.

In the demo, the researchers put a glove on the robotic hand, tied up some of its fingers, and even used a plush giraffe to interfere with its movements. The robotic hand would make mistakes and needed adjustments, but it could still keep going.

There is an easily overlooked boundary to this achievement: how to solve the cube was computed by another algorithm; what Dactyl had to solve was how to execute those steps with a robotic hand. Even so, getting a robotic hand to complete such a complex manipulation task in the real world was still a notable advance at the time.

But this did not make robotics continue to be a focus of OpenAI's research.

Key events for OpenAI from 2016 to 2026

In 2020, OpenAI shut down this line of research. It was not until 2021, when co-founder Wojciech Zaremba discussed the matter on a podcast, that the outside world learned more clearly that the team had been disbanded.

His explanation was data. For a robot to learn an action, it needs hardware, space, and experiments. Equipment breaks, data collection is slow, and if you switch to a different robotic hand or a different environment, the previous experience may not transfer directly. By contrast, the internet already holds vast amounts of text that can be used continuously to train language models.

For a company aiming at artificial general intelligence, the research efficiency of the two directions began to diverge. In 2020, GPT-3 was released, and language models had already shown the possibility of gaining more capabilities as training scale increased. OpenAI concentrated its resources on this path.

Over the next two years, the robotics industry kept moving forward. Tesla announced the Optimus plan in 2021 and showed a prototype in 2022. Also in 2022, OpenAI launched ChatGPT.

ChatGPT's success quickly changed the way the robotics industry discussed its problems. Since a model can acquire abilities such as conversation, coding, and reasoning by learning from text, could robots also have a general-purpose "brain" and not need to be programmed separately for each task?

OpenAI once again appeared on the funding lists of robotics companies.

In March 2023, it led a funding round for the Norwegian robotics company 1X. In February 2024, it participated in Figure's $675 million Series B financing and signed a cooperation agreement with the company to develop models for humanoid robots.

This time, OpenAI did not build the hardware itself but partnered with companies that had robot bodies. A demo released by Figure in March 2024 made this division of labor intuitive: the robot could answer people's questions while identifying objects on the table, handing over an apple, and clearing the dishes. OpenAI provided model capabilities, and Figure was responsible for connecting them to a real body.

But the partnership did not last long.

In February 2025, Figure announced the end of its cooperation with OpenAI, shifted to developing its own models, and subsequently released Helix. Figure CEO Brett Adcock emphasized that robot AI needs vertical integration with hardware.

This exposed a difficult problem in the partnership model. Model companies can provide language and vision capabilities, but how a robot moves every day, where it fails, and how it completes tasks—these experiences are generated on the hardware side. Who collects the data and how it is used for the next round of training require close coordination between both sides. For Figure, controlling its own models could shorten this process.

Meanwhile, OpenAI was already rebuilding its own robotics business.

In 2024, news gradually emerged that it was hiring robotics researchers. In November of the same year, Caitlin Kalinowski, who had worked on Apple hardware and the Meta Orion glasses, joined OpenAI to lead robotics and consumer hardware.

By early 2025, job postings began to involve specific roles in sensors and mechanical structures. OpenAI's plans gradually extended beyond model research to robot bodies. Subsequent media reports also mentioned that it had set up a robotics testing site in San Francisco to collect real motion data through teleoperation.

The first time OpenAI did robotics, it tried to use simulation to make up for the shortage of real-world data. After re-entering the field, it began investing people and equipment to collect data itself.

The industry is also advancing in a similar direction. In 2025, Google DeepMind released Gemini Robotics, and Nvidia launched GR00T N1. Robot manufacturers, model companies, and chip companies are all expanding their investments, trying to connect vision and language understanding with motion control.

However, OpenAI's internal robotics business has always been in a state of flux.

In March 2026, Kalinowski resigned due to controversy over OpenAI's cooperation with the Pentagon. On May 31, Altman publicly announced that OpenAI Robotics was hiring. In his post, he said that the world simulation research led by Aditya Ramesh had evolved into this team over the past year. This was also the first time the Robotics team made a public appearance.

At the time, the goals Altman gave were very specific: in the short term, build robots that can help skilled workers build infrastructure; in the long term, he hopes everyone can own a personal robot.

In the months that followed, the business still appeared to be expanding. Altman confirmed that OpenAI would develop humanoid robots and proposed a plan to hold a demonstration in 2027. On the recruitment page, software, control, mechanical, and hardware positions were still continuously hiring.

Until November 13, Fortune reported that OpenAI had dissolved the independent robotics team and reassigned its personnel to other departments. It had been 166 days since Altman publicly recruited.

Researchers were transferred into the core research system, and applied robotics and hardware personnel entered the Applied department. The independent team was no longer retained, while the related work continued to exist within the company.

This differs from how it was handled in 2020. That year, OpenAI stopped robotics research and shifted resources to language models, which made it easier to obtain data and easier to scale training. This time, the reason is that general models can already take on part of the robotics tasks, and robotics remains the company's long-term direction.

Between the two adjustments, OpenAI also tried investment and cooperation to have external companies provide hardware.

So over these ten years, OpenAI has been constantly changing the way it participates in robotics, such as which work it does itself, which it hands to partners, and which is worth keeping an independent team for.

Of course, we still do not know the full process of OpenAI's internal decision-making. But one month before the news was announced, a study provided a noteworthy backdrop: general models had already begun to achieve results in robot manipulation tests that surpassed specialized policies.

On October 14, Nvidia researchers released Astranomy, testing whether GPT-6 Astra could complete manipulation tasks in a simulated environment without additional robot policy fine-tuning. On the RoboLab-120 benchmark, the best-performing agent configuration achieved a 74.8% success rate, higher than the 51% of the specialized policy Lakeside on the leaderboard. However, this result depended on a Codex configuration that continuously retained context; directly calling the model to generate actions yielded a 66.3% success rate.

Just at the beginning of 2026, the internal roadmap of the robotics team still assumed that to give general models manipulation capabilities, it would first be necessary to use a Sora-style video model to accumulate physical intuition, and then distill it into robot policies. In less than a year, this intermediate step was skipped.

In this regard, Carnegie Mellon University professor Guanya Shi interpreted OpenAI's disbanding of the robotics team as follows: in OpenAI's view, robotics is no longer a special problem; it is merely a downstream application of general intelligence.

In other words, judged from a purely technical perspective, this is not a retreat but an absorption, putting the exploration of robotics back into the main mission.

But any decision by OpenAI has never been based on technology alone.

Another more realistic backdrop is money. According to media reports, OpenAI's adjusted operating loss in the third quarter of 2026 reached $10.9 billion. The company is also preparing for a possible IPO. Revenue is growing rapidly, and spending continues to expand, so management must constantly decide which businesses are worth separately investing personnel and computing power in.

In March of this year, Fidji Simo, head of the applied business, had already told employees to focus on core businesses and avoid being distracted by side projects. The subsequent contraction of Sora also occurred against this backdrop.

In other words, it is hard to understand the disbanding of the robotics team entirely apart from the pressure of resource allocation.

Technological progress and business pressure may have jointly driven this decision. General models beginning to show manipulation capabilities gave OpenAI reason to reassess the necessity of independent investment; continuously expanding spending made that assessment more urgent.

OpenAI's past success makes it easy to attribute some kind of foresight to this adjustment as well. Leaving robotics in 2020 and concentrating on language models did indeed bring huge returns later. But that outcome cannot prove today's choice in advance. A 74.8% simulation success rate is still a long way from a robot that can work stably at controllable cost.

In this regard, after Tiger Sniff spoke with many practitioners in the embodied intelligence industry, they all said that today's GPT only shows us that large language models are helpful for solving problems in the embodied industry, but "helpful" does not mean it can replace embodied models; it is still very far from the goal.

An important window for observation is 2027. Altman previously said that OpenAI plans to showcase its self-developed robot that year.

As for whether OpenAI has found a faster path or reduced investment too early, the answer still lies outside the laboratory.

Article title: OpenAI disbands robotics team

Published 96 articles on Tiger Sniff. This content may not be reproduced without permission. For licensing matters, please contact hezuo@huxiu.com. Everyone who is changing and wants to change the world is on the Tiger Sniff app[ ]

Source

Huxiu