# OpenAI Robotics

> Source: https://aiwiki.ai/wiki/openai_robotics
> Updated: 2026-09-23
> Fact-checked: 2026-09-23
> Categories: Embodied AI, Humanoid Robots, OpenAI, Robotics
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "OpenAI Robotics." aiwiki.ai, 23 Sept 2026. https://aiwiki.ai/wiki/openai_robotics
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**OpenAI Robotics** is the robotics research and hardware group at [OpenAI](https://aiwiki.ai/wiki/openai). OpenAI has worked on robots in two separate periods. From 2016 to 2020 a research team trained robot controllers in simulation with [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning); its best-known results were the [Dactyl](https://aiwiki.ai/wiki/dactyl) robot hand (2018) and a one-handed Rubik's Cube solve (2019). OpenAI decided in October 2020 to stop that work, and the decision became public in July 2021. After investing in outside robotics startups from 2023, OpenAI started hiring for a new internal robotics team in 2024. On 31 May 2026 chief executive [Sam Altman](https://aiwiki.ai/wiki/sam_altman) said that the company's world simulation research program, led by [Aditya Ramesh](https://aiwiki.ai/wiki/aditya_ramesh), had "evolved over the past year into OpenAI Robotics."[27] By September 2026 the group was hiring actuator, motor, firmware and data-collection staff, and Altman had said OpenAI would "definitely" build a [humanoid robot](https://aiwiki.ai/wiki/humanoid_robot).[1][30]

## At a glance

| Period | What OpenAI did | Key sources |
|---|---|---|
| June 2016 | Listed "Build a household robot" as one of four technical goals, using an off-the-shelf robot | [3] |
| 2017 | Published sim-to-real work: domain randomization, one-shot imitation, dynamics randomization | [4][5][6] |
| February 2018 | Released eight simulated robotics environments for Gym and a Hindsight Experience Replay implementation | [7][8] |
| July 2018 | Dactyl: a Shadow Dexterous Hand reorients a block after training only in simulation | [9][10] |
| October 2019 | The same hand solves a Rubik's Cube, trained with Automatic Domain Randomization | [11][12] |
| October 2020 | Decides to stop robotics research (disclosed July 2021) | [13] |
| 2023-2024 | OpenAI Startup Fund invests in 1X and Figure; OpenAI invests in Physical Intelligence; Figure collaboration agreement | [15][16][17] |
| 2024 | Begins hiring research engineers for a rebooted robotics team; Caitlin Kalinowski joins in November | [14][19] |
| January 2025 | Job listings and a post by Kalinowski say OpenAI will build its own robots with custom sensors | [20] |
| February 2025 | Figure exits its OpenAI collaboration; OpenAI starts a data-collection lab in San Francisco | [18][23] |
| March 2026 | Kalinowski resigns; OpenAI discontinues the Sora app and says the Sora research team continues to focus on world simulation research for robotics | [25][26] |
| May 2026 | Altman announces OpenAI Robotics hiring, led by Aditya Ramesh | [27] |
| September 2026 | Business Insider counts 27 robotics job listings, with base salaries up to $500,000 | [1] |

## Early robotics research (2016-2019)

### The household robot goal

In a post dated 20 June 2016, OpenAI's founders set out four technical goals. Goal 2 was "Build a household robot." The post said OpenAI was working "to enable a physical robot (off-the-shelf; not manufactured by OpenAI) to perform basic housework."[3] The same post described a planned metric built from [Gym](https://aiwiki.ai/wiki/gym) environments spanning games, robotics and language tasks.[3] Cofounder [Wojciech Zaremba](https://aiwiki.ai/wiki/wojciech_zaremba) oversaw the team that pursued this "general purpose robot."[14]

### Training in simulation

OpenAI's robotics work in 2017 centered on training in simulation and then running the result on a real robot. A March 2017 paper led by Josh Tobin, "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World," trained an object detector only on simulated images with non-realistic random textures. The detector was accurate to 1.5 cm on real images and was used for grasping in clutter.[4] On 16 May 2017 OpenAI described a system, "trained entirely in simulation and deployed on a physical robot," that could learn a new block-stacking task after seeing a human perform it once in virtual reality. It combined a vision network trained on simulated images with a one-shot [imitation learning](https://aiwiki.ai/wiki/imitation_learning) network. The vision system "is never trained on a real image."[5] That month OpenAI also released Roboschool, open-source software for robot control in simulation.[13]

In October 2017 OpenAI introduced dynamics randomization. During training it randomized 95 properties of the simulated environment, such as link masses, friction, table height, action latency and observation noise. It then trained an [LSTM](https://aiwiki.ai/wiki/lstm)-based policy that pushed a puck to a goal on a real robot arm, even when the puck sat on a bag of chips.[6] The post said OpenAI saw three routes to general-purpose robots: huge fleets of physical robots, simulators that closely match reality, and randomized simulators. It added: "We increasingly believe that the third will be the most important part of the solution."[6] The effort became known as [sim-to-real transfer](https://aiwiki.ai/wiki/sim_to_real_transfer).

### Gym robotics environments

On 26 February 2018 OpenAI released eight simulated robotics environments for Gym, built on the [MuJoCo](https://aiwiki.ai/wiki/mujoco) physics simulator. Four used the Fetch research robot and four used the Shadow Dexterous Hand. All were goal-based tasks with sparse rewards by default. The release included a Baselines implementation of Hindsight Experience Replay (HER), an algorithm that treats a failed attempt as success at a different goal.[7] A technical report, "Multi-Goal Reinforcement Learning: Challenging Robotics Environments and Request for Research," came out the same day.[8]

### Dactyl (2018)

On 30 July 2018 OpenAI announced Dactyl, a system that trained a Shadow Dexterous Hand to rotate a block or prism in its palm into a target orientation. Dactyl was "trained entirely in simulation" and transferred to the real hand without fine-tuning. It used the same general-purpose reinforcement learning algorithm and code as [OpenAI Five](https://aiwiki.ai/wiki/openai_five), OpenAI's Dota 2 system.[9] Training used Rapid, OpenAI's large-scale implementation of [Proximal Policy Optimization](https://aiwiki.ai/wiki/ppo), on 6,144 CPU cores and 8 GPUs, and collected "about one hundred years of experience in 50 hours."[9] A convolutional network estimated the object's pose from three ordinary RGB cameras and was also trained only on simulated images.[9]

With full randomization, the policy that read pose from motion-capture tracking achieved a median of 13 consecutive rotations on the real hand (maximum 50). Using camera-based pose estimation it achieved a median of 11.5 (maximum 46). A policy trained without randomization reached a median of 0.[9] OpenAI reported that learning the task without randomization took about 3 years of simulated experience, and reaching similar performance under full randomization took about 100 years.[9] The paper, "Learning Dexterous In-Hand Manipulation," was posted to arXiv on 1 August 2018.[10]

### Rubik's Cube (2019)

On 15 October 2019 OpenAI published "Solving Rubik's Cube with a robot hand." The post said the team had been trying to get a humanlike hand to solve the cube since May 2017, had solved it in simulation in July 2017, and as of July 2018 could only manipulate a block on the real robot.[11] The move sequence came from Kociemba's algorithm. The neural networks performed the face rotations and cube flips.[11]

The main new method was Automatic Domain Randomization (ADR). ADR starts from a single non-randomized environment and widens the randomization ranges, such as cube size, cube mass and finger friction, each time the network passes a performance threshold.[11] OpenAI reported that the resulting system handled perturbations it never saw in training, including a rubber glove, tied fingers, a blanket and prodding by a plush giraffe. Its stated hypothesis was that memory-augmented networks combined with a sufficiently randomized environment lead to "emergent meta-learning."[11] The hand solved the cube 20% of the time from a maximally difficult scramble needing 26 face rotations, and 60% of the time from scrambles needing 15.[11] The paper was posted to arXiv on 16 October 2019.[12]

| Project | Date | Hardware | Method | Reported result |
|---|---|---|---|---|
| Domain randomization for object detection | March 2017 | Robot arm with camera | Detector trained on randomized simulated images | Real-world detector accurate to 1.5 cm, trained without real images [4] |
| One-shot imitation ("Robots that learn") | May 2017 | Robot arm | Simulated vision network plus imitation network | Block stacking from a single VR demonstration [5] |
| Dynamics randomization | October 2017 | Robot arm | 95 randomized dynamics properties, LSTM policy | Puck pushing robust to changed surfaces [6] |
| Gym robotics environments | February 2018 | Fetch, Shadow Hand (simulated) | Sparse-reward goal tasks, HER | Eight environments released [7] |
| Dactyl | July 2018 | Shadow Dexterous Hand | PPO (Rapid) with domain randomization | Median 13 block rotations (motion tracking) [9] |
| Rubik's Cube | October 2019 | Shadow Dexterous Hand | PPO with Automatic Domain Randomization | 20% solve rate at 26 rotations, 60% at 15 [11] |

## Shutdown of the first team (2020-2021)

The first robotics team wound down in 2020, but this was not public until mid-2021, which is why sources give both years. On 16 July 2021 VentureBeat reported that Zaremba had disclosed on a podcast hosted by Weights & Biases that the team was disbanded. An OpenAI spokesperson told VentureBeat that "last October we decided not to pursue further robotics research and instead refocus the team on other projects." The spokesperson added that "other approaches, such as reinforcement learning with human feedback, lead to faster progress in our reinforcement learning research."[13] Forbes and Business Insider later dated the closure to October 2020 and 2020 respectively. Wired and other outlets have used 2021, the year it was announced.[14][23][22]

Zaremba explained the decision in terms of data: "So it turns out that we can make a gigantic progress whenever we have access to data... And ultimately that was holding us back in terms of robotics." He also said: "The sad thing is, if we were a robotics company, the mission of the company would be different, and I think we would continue."[13]

Several members of the robotics team stayed at OpenAI in other roles. By 2024 Zaremba was helping lead GPT development, Peter Welinder led product and partnerships, [Bob McGrew](https://aiwiki.ai/wiki/bob_mcgrew) was vice president of research, and [Lilian Weng](https://aiwiki.ai/wiki/lilian_weng) was head of safety systems.[14] Forbes also noted that [Covariant](https://aiwiki.ai/wiki/covariant) had been started by former OpenAI robotics team members.[14]

## Investing in outside robot makers (2023-2025)

With no in-house team, OpenAI took part in robotics through investments, mostly by the OpenAI Startup Fund.

| Company | Date | Round | OpenAI role | Source |
|---|---|---|---|---|
| [1X Technologies](https://aiwiki.ai/wiki/1x_technologies) | 23 March 2023 | $23.5 million Series A2 | OpenAI Startup Fund led the round | [15] |
| [Figure AI](https://aiwiki.ai/wiki/figure_ai) | 29 February 2024 | $675 million Series B at a $2.6 billion valuation | OpenAI Startup Fund among investors; separate collaboration agreement | [16] |
| [Physical Intelligence](https://aiwiki.ai/wiki/physical_intelligence) | 2024 | $70 million raised by May 2024; $400 million round at a $2.4 billion post-money valuation (4 November 2024) | Investor | [14][17] |

In the 1X announcement, [Brad Lightcap](https://aiwiki.ai/wiki/brad_lightcap), OpenAI's chief operating officer and manager of the Startup Fund, said the fund "believes in the approach and impact that 1X can have on the future of work."[15]

The Figure deal went further than an investment. Figure and OpenAI signed "a collaboration agreement to develop next generation AI models for humanoid robots." Welinder, then OpenAI's vice president of product and partnerships, said: "We've always planned to come back to robotics and we see a path with Figure to explore what humanoid robots can achieve when powered by highly capable multimodal models."[16] A month later Figure released a video of its robot talking and reasoning with support from a multimodal model trained by OpenAI.[14] In August 2024 the two companies said [Figure 02](https://aiwiki.ai/wiki/figure_02) would use OpenAI models for natural-language communication.[18]

On 4 February 2025 Figure said on X that it was leaving the agreement, citing a "major breakthrough" in its own AI. Founder [Brett Adcock](https://aiwiki.ai/wiki/brett_adcock) told TechCrunch: "We found that to solve embodied AI at scale in the real world, you have to vertically integrate robot AI... We can't outsource AI for the same reason we can't outsource our hardware."[18] In September 2026 Business Insider still described the OpenAI Startup Fund as an investor in Figure.[1]

## Rebuilding an internal team (2024-2025)

### The 2024 reboot

On 30 May 2024 Forbes reported that OpenAI was hiring research engineers to rebuild its robotics team. A job listing said new hires would be "one of the first members of the team," and a source said the group had existed for about two months. After publication, OpenAI confirmed it had begun hiring for the team.[14] Two sources told Forbes that OpenAI then meant to "coexist rather than compete" with robot makers, supplying technology that they would build into their own systems. Forbes's sources said it was unclear whether OpenAI would develop robot hardware.[14]

On 4 November 2024 Caitlin Kalinowski, who had led Meta's augmented reality glasses team and the hardware for its [Orion](https://aiwiki.ai/wiki/meta_orion) prototype, announced that she was "joining OpenAI to lead robotics and consumer hardware." She wrote that she would "initially focus on OpenAI's robotics work and partnerships to help bring AI into the physical world."[19] In December 2024 The Information reported that OpenAI was considering developing its own humanoid robots.[22]

### Own hardware, custom sensors

In January 2025 OpenAI's plans shifted toward building hardware. In a post on X on 10 January 2025, Kalinowski said OpenAI would develop its own robots with a custom sensor suite. The job listings published with it described "general-purpose," "adaptive" and "versatile" robots, eventual "full-scale production," and a requirement for "experience designing mechanical systems intended for high volume (1M+)."[20] A trademark application OpenAI filed with the U.S. Patent and Trademark Office in late January 2025 covered "user-programmable humanoid robots" and "humanoid robots having communication and learning functions for assisting and entertaining people." TechCrunch noted that such applications are often written broadly.[21]

### The San Francisco data-collection lab

Business Insider reported in January 2026 that OpenAI had "quietly built up a humanoid robotics lab" over the previous year. According to insiders, the lab launched in February 2025, shared a San Francisco building with OpenAI's finance team, employed about 100 data collectors, and had more than quadrupled in size.[23] Most of the work was [teleoperation](https://aiwiki.ai/wiki/teleoperation). Workers used 3D-printed controllers called GELLOs to operate two Franka robot arms. The first task was putting a rubber duck in a cup; later tasks included putting bread in a toaster and folding laundry.[23] The lab ran three shifts around the clock, and workers were rated on the "good hours" of usable training data they produced. A humanoid robot described as "iRobot-like" was on display but was mostly unused.[23] In December 2025 OpenAI told employees it planned a second lab in Richmond, California.[23]

Business Insider contrasted this with the first program. The earlier work relied on reinforcement learning, while the new effort collected large amounts of human demonstration data.[23] Wired reported in September 2025 that OpenAI was recruiting researchers who worked on humanoid systems, including Chengshu Li, who joined from Stanford in June 2025. One job opening asked for expertise in teleoperation and simulation tools including [NVIDIA](https://aiwiki.ai/wiki/nvidia) Isaac.[22]

## OpenAI Robotics (2026)

### Supply chain, departures and the end of the Sora app

On 15 January 2026 OpenAI issued a request for proposals for U.S.-based manufacturing covering data-center hardware, consumer electronics and robotics. For robotics it sought "critical inputs for advanced robotics (e.g., gearboxes, motors, power electronics)."[24]

On 7 March 2026 Kalinowski resigned in response to OpenAI's agreement with the U.S. Department of Defense. She wrote that "surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got," and that her decision was "about principle, not people."[25]

On 24 March 2026 OpenAI said it would discontinue its Sora video app. A spokesperson said that "the Sora research team continues to focus on world simulation research to advance robotics that will help people solve real-world, physical tasks."[26]

### The May 2026 announcement

On 31 May 2026 Altman posted that "OpenAI Robotics is hiring," seeking "full-stack hardware, ops, systems, and ML engineers to help us program and manufacture robots that are useful for society." He set out two horizons: "In the short term, we are focused on robots to support skilled workers to build our future infrastructure; in the long term, we imagine everyone having a personal robot doing anything they need." He wrote that the world simulation research program led by Aditya Ramesh "has evolved over the past year into OpenAI Robotics," and that progress rested on "co-design between robotics hardware and ML research."[27] Ramesh, who created the original [DALL-E](https://aiwiki.ai/wiki/dall_e) and led work on [Sora](https://aiwiki.ai/wiki/sora), describes himself as leading Worldsim as a VP of Research and "bootstrapping our new robotics effort to bring the intelligence of our video generation models to the physical world."[28]

In July 2026 CnEVPost, citing Leiphone, reported that Lu Siyuan, head of [XPeng](https://aiwiki.ai/wiki/xpeng)'s AI infrastructure department, was leaving to work on embodied AI robotics at OpenAI.[29]

### "We will definitely do a humanoid"

In a TIME article published on 26 August 2026, Alex Heath reported that Altman said OpenAI will "definitely" make humanoid robots, and that "someday, Altman believes, everyone should have a personal robot."[30] Business Insider quoted Altman from Heath's Sources podcast: "We will definitely do a humanoid. We will do other form factors as well." He also said: "There will of course be data center robots that have, like, different form factors."[1]

## September 2026 hiring

On 18 September 2026 Business Insider reported that OpenAI had 27 robotics jobs on its careers page, up from 11 in May, covering hardware, software, data collection and prototyping. It said advertised base salaries ranged from $177,000 to $500,000 a year before equity.[1] The listings included actuator design engineers, software, firmware and machine-learning engineers, a lab technician to "develop, build, test, and iterate on robotic systems," a lawyer for the robotics team, and a manager for OpenAI's "data collection facilities." The highest-paid opening was a machine-learning engineer for distributed data systems, at up to $500,000.[1] Business Insider reported that Ramesh leads the robotics team. OpenAI declined to comment.[1]

Guy Hoffman, a Cornell professor who reviewed all 27 postings for Business Insider, said OpenAI appeared to be assembling a "custom robot design team." He said references to laser range finders and batteries "vaguely imply" an untethered mobile robot, without showing whether it would use wheels or legs, and that the small amount of electronics work suggested little focus on sensors "beyond off-the-shelf cameras, microphones, and laser range finders."[1]

### Snapshot of the job board

OpenAI's public job board, read on 23 September 2026, listed 22 openings under the Robotics team. Two more robotics-specific roles sat under other teams: a senior commercial counsel for robotics and a senior environmental health and safety manager for robotics. All 24 were based in San Francisco. Advertised base salary ranges ran from $157,000 (technical commodity manager) to $500,000.[2] Counts and ranges shift as postings open and close, so this snapshot differs from the Business Insider figures.

| Role (selection) | Team | Base salary range (USD) |
|---|---|---|
| Machine Learning Engineer, Distributed Data Systems - Robotics | Robotics | $380,000-$500,000 |
| Lead Thermal Simulation Engineer | Robotics | $423,000-$485,000 |
| PCB Layout Engineer, Robotics | Robotics | $318,000-$485,000 |
| Control Systems Software Engineer, Robotics | Robotics | $380,000-$460,000 |
| Firmware Engineer, Robotics | Robotics | $380,000-$445,000 |
| Actuator Design Engineer | Robotics | $225,000-$423,000 |
| Manufacturing Engineer, Motors & Actuators | Robotics | $225,000-$318,000 |
| Prototyping Lab Technician, Robotics | Robotics | $225,000-$295,000 |
| Operations Program Manager - Robotics Data Acquisition | Robotics | $177,000-$251,000 |
| Technical Commodity Manager - Robotics | Robotics | $157,000-$196,000 |
| Senior Counsel, Commercial (Robotics) | Legal | $297,000-$347,000 |

Most of the postings share a team description: the Robotics team "is focused on unlocking general-purpose robotics and pushing towards AGI-level intelligence in dynamic, real-world settings" and explores "a broad range of robotic form factors."[2] The same language appeared in listings Wired reviewed in 2025.[22] The firmware posting repeats Altman's framing of robots for skilled infrastructure workers in the short term and "everyone having a personal robot doing anything they need" in the long term.[2] The data-acquisition role describes running OpenAI's "data collection facilities," including workstations, operators, rigs and data quality.[2] Seven of the 24 listings concern actuators, motors, gears or their test infrastructure.[2]

## Then and now

| Aspect | First team (2016-2020) | OpenAI Robotics (2024-2026) |
|---|---|---|
| Hardware | Off-the-shelf robots (Shadow Dexterous Hand, Fetch, research arms); the 2016 goal specified a robot "not manufactured by OpenAI" [3][7] | In-house design of actuators, motors and prototypes, with plans for volume manufacturing [2][20][24] |
| Training data | Simulated experience with domain randomization [9][11] | Human teleoperation data from dedicated collection facilities, plus simulation [23][2] |
| Research lineage | Reinforcement learning shared with OpenAI Five [9] | World simulation research that grew out of the Sora video models [27][28] |
| Stated goal | A household robot [3] | Robots for skilled infrastructure workers first, personal robots later [27] |
| Leadership | Wojciech Zaremba [14] | Caitlin Kalinowski (2024-2026); Aditya Ramesh [19][25][27][1] |

## References

1. Rya Jetha. "OpenAI is offering engineers up to $500,000 in salary as it pushes into robotics." Business Insider, 18 September 2026. https://www.businessinsider.com/openai-is-offering-robotics-engineers-up-to-500-000-in-salary-2026-9
2. OpenAI job board (Ashby), robotics postings, accessed 23 September 2026. https://jobs.ashbyhq.com/openai
3. OpenAI. "OpenAI technical goals." 20 June 2016. https://openai.com/index/openai-technical-goals/
4. Josh Tobin et al. "Domain Randomization for Transferring Deep Neural Networks from Simulation to the Real World." arXiv:1703.06907, March 2017. https://arxiv.org/abs/1703.06907
5. OpenAI. "Robots that learn." 16 May 2017. https://openai.com/index/robots-that-learn/
6. OpenAI. "Generalizing from simulation." 19 October 2017. https://openai.com/index/generalizing-from-simulation/
7. OpenAI. "Ingredients for robotics research." 26 February 2018. https://openai.com/index/ingredients-for-robotics-research/
8. Matthias Plappert et al. "Multi-Goal Reinforcement Learning: Challenging Robotics Environments and Request for Research." arXiv:1802.09464, February 2018. https://arxiv.org/abs/1802.09464
9. OpenAI. "Learning dexterity." 30 July 2018. https://openai.com/index/learning-dexterity/
10. OpenAI et al. "Learning Dexterous In-Hand Manipulation." arXiv:1808.00177, August 2018. https://arxiv.org/abs/1808.00177
11. OpenAI. "Solving Rubik's Cube with a robot hand." 15 October 2019. https://openai.com/index/solving-rubiks-cube/
12. OpenAI et al. "Solving Rubik's Cube with a Robot Hand." arXiv:1910.07113, October 2019. https://arxiv.org/abs/1910.07113
13. "OpenAI disbands its robotics research team." VentureBeat, 16 July 2021. https://venturebeat.com/business/openai-disbands-its-robotics-research-team
14. Kenrick Cai. "OpenAI Is Rebooting Its Robotics Team." Forbes, 30 May 2024. https://www.forbes.com/sites/kenrickcai/2024/05/30/openai-robotics-team/
15. 1X Technologies. "1X Raises $23.5M in Series A2 Funding led by OpenAI." 23 March 2023. https://www.1x.tech/discover/1x-rasies-23-5m-in-series-a2-funding-led-by-open-ai
16. Figure. "Figure Raises $675M at $2.6B Valuation and Signs Collaboration Agreement with OpenAI." PR Newswire, 29 February 2024. https://www.prnewswire.com/news-releases/figure-raises-675m-at-2-6b-valuation-and-signs-collaboration-agreement-with-openai-302074897.html
17. "Jeff Bezos and OpenAI invest in robot startup Physical Intelligence at $2.4 billion valuation." CNBC, 4 November 2024. https://www.cnbc.com/2024/11/04/jeff-bezos-and-openai-invest-in-robot-startup-physical-intelligence.html
18. "Figure drops OpenAI in favor of in-house models." TechCrunch, 4 February 2025. https://techcrunch.com/2025/02/04/figure-drops-openai-in-favor-of-in-house-models/
19. "Meta's former hardware lead for Orion is joining OpenAI." TechCrunch, 4 November 2024. https://techcrunch.com/2024/11/04/metas-former-hardware-lead-for-orion-is-joining-openai/
20. Kyle Wiggers. "New OpenAI job listings reveal the company's robotics plans." TechCrunch, 10 January 2025. https://techcrunch.com/2025/01/10/new-openai-job-listings-reveal-its-robotics-plans/
21. "OpenAI's new trademark application hints at humanoid robots, smart jewelry, and more." TechCrunch, 3 February 2025. https://techcrunch.com/2025/02/03/openais-new-trademark-application-hints-at-humanoid-robots-smart-jewelry-and-more/
22. Will Knight. "OpenAI Ramps Up Robotics Work in Race Toward AGI." Wired, 15 September 2025. https://www.wired.com/story/openai-ramps-up-robotics-work-in-race-toward-agi/
23. Grace Kay. "Inside OpenAI's renewed push into robotics." Business Insider, 22 January 2026. https://www.businessinsider.com/open-ai-robotics-lab-humanoid-robots-2026-1
24. OpenAI. "Strengthening the US AI supply chain through domestic manufacturing." January 2026. https://openai.com/index/strengthening-the-us-ai-supply-chain/
25. "OpenAI hardware exec Caitlin Kalinowski quits in response to Pentagon deal." TechCrunch, 7 March 2026. https://techcrunch.com/2026/03/07/openai-robotics-lead-caitlin-kalinowski-quits-in-response-to-pentagon-deal/
26. "OpenAI is shutting down its Sora video app just months after launch." CNN, 24 March 2026. https://edition.cnn.com/2026/03/24/tech/openai-sora-video-app-shutting-down
27. Sam Altman. Post on X, 31 May 2026. https://x.com/sama/status/2061117302528188712
28. Aditya Ramesh. Personal website, accessed 23 September 2026. http://adityaramesh.com/
29. "Xpeng loses AI infrastructure chief to OpenAI's robotics push, report says." CnEVPost, 20 July 2026. https://cnevpost.com/2026/07/20/xpeng-loses-ai-infra-chief-to-openai/
30. Alex Heath. "Inside OpenAI's Reboot." TIME, 26 August 2026. https://time.com/article/2026/08/26/openai-sam-altman-interview/