OpenAI launches its robotics division: from software AGI to embodied intelligence
🔎 Why is the software giant getting into metal?
On May 31, 2026, Sam Altman posts a simple message on X: OpenAI is hiring for its new robotics division. Behind this laconic announcement, the signal is blunt. The world's most powerful software AI company decides to build physical robots.
This pivot is not a whim. It stems directly from the internal world simulation program, these models capable of predicting and generating 3D environments. When you understand the world well enough to simulate it, the next logical step is to act within it. This is exactly what OpenAI is doing.
The strategic context adds urgency. Tesla is deploying its Optimus robots in factories, Figure AI lance Helix, Genesis AI dévoile GENE-26.5 avec des mains robotiques humanoïdes. OpenAI risked becoming the brain supplier for other people's robots. Altman decided to no longer be just a supplier.
The essentials
- OpenAI is creating an official robotics division, announced on May 31, 2026 by Sam Altman on X, with open recruitment for full-stack hardware, ops, systems, and ML engineers.
- The division is led by Aditya Ramesh, co-creator of DALL-E, and stems directly from OpenAI's world simulation program — the logic is clear: simulate the world, then physically act within it.
- The stated goal is a mass-market personal robot ("imagine everyone having a personal robot"), which puts OpenAI in direct confrontation with Elon Musk's Tesla Optimus.
- OpenAI is starting without a clear hardware ally after the end of its partnership with Figure AI in February 2025, unlike Tesla's vertically integrated approach.
- The central question becomes: which LLM will serve as the brain for these robots, and how will the reasoning capabilities of GPT-5.5 or Claude Opus 4.7 be adapted for real-time?
Recommended tools
| Tool | Main use | Price (June 2026, check on openai.com) | Ideal for |
|---|---|---|---|
| GPT-5.5 | Agentic LLM, complex reasoning | Pro/Team subscription | Robot brain, task planning |
| Gemini 3 Pro Deep Think | Deep multi-step reasoning | Google AI subscription | Complex environment simulation |
| Claude Opus 4.7 (Adaptive) | Long-context adaptive reasoning | Pro subscription | Long spatial scene analysis |
| Hostinger | Web hosting for robotics projects | Starting at 2.99 €/month | Deployment of control dashboards |
The genesis: from world simulation to physical robot
OpenAI's robotics division does not come out of nowhere. It is the direct extension of internal work on world simulation (world models). According to the detailed analysis by AIMind Pub, this program aimed to create models capable of understanding physics, geometry, and causality in 3D environments.
A world model learns to predict what happens when an object falls, when a door opens, when a liquid is poured. This is fundamentally the same skill a robot needs to navigate an apartment. Simulation is the training. The robot is the deployment.
What is striking is the strategic continuity. OpenAI first learned to generate text (GPT), then images (DALL-E, led by Ramesh), then to simulate worlds. Each step adds a layer of understanding of reality. The robot is the culmination of this chain.
AIBase News emphasizes that this transition from software to hardware marks a turning point in the company's history. Until now, OpenAI excelled in a domain where errors have no physical consequences. A robot that makes a mistake can break an object, injure someone. The level of requirement changes dimension.
Aditya Ramesh : why the co-creator of DALL-E is leading the robotics division
Ramesh's appointment is not anecdotal. It is a strong strategic signal. According to the Economic Times, he is the one leading this new division.
Ramesh is not a career roboticist. He is a generative AI researcher. He led DALL-E, the model that proved an LLM could understand and generate visual reality. Moving from generated images to physical robots follows the exact same logic: understanding the world to produce it.
It is also an internal political choice. Ramesh is one of OpenAI's most prestigious talents, of Indian origin, which India Today highlights. Entrusting him with this project signals that it is an absolute priority, not a side project.
The profile sought for the team, as reported by Clanks, is revealing: "exceptional full-stack hardware, ops, systems, and ML engineers". The term full-stack applied to hardware means that OpenAI wants to control the entire chain, from sensor to actuator, from model to metal.
The robot's brain: which LLM for embodied intelligence?
It's the question no one dares to clearly formulate yet. A humanoid robot needs a model capable of three things simultaneously: perceiving (vision + sensors), planning (sequential reasoning), and acting (real-time generation of motor commands).
GPT-5.5, currently the top agentic LLM with a score of 98.2, seems like the natural candidate. Its multi-step planning capability is exactly what is needed to break down "making coffee" into a sequence of physical actions. But a text-based LLM is not designed for real-time.
Several approaches are possible. The first: a main model like GPT-5.5 for high-level planning, coupled with a lighter and faster model for closed-loop motor control. The second: using a model like Anthropic's Claude Opus 4.7 (Adaptive), whose adaptive mode could theoretically adjust its level of reasoning based on the urgency of the physical situation.
The third approach, the most likely, is the one implicitly described by Comparos: OpenAI will develop specialized models for robotics, trained on its internal world models, which will not simply be generalist LLMs plugged into servomotors.
What is certain is that reasoning models like Gemini 3 Pro Deep Think (agentic score: 95.4) or o1-preview (90.2) show that the ability to chain logical deductions has become mature enough to consider a transfer to the physical domain. The question is no longer if an LLM can pilot a robot, but what level of latency is acceptable.
OpenAI vs Tesla vs Figure AI : a three-way race with uneven starting lines
Tesla : the overwhelming advantage of vertical integration
Tesla Optimus is already in the factory. This is the point that Gagadget rightly hammers home. Tesla controls the entire chain: motors, sensors, production factory, software, data collected by millions of vehicles. It's a monstrous structural advantage.
Musk started working on the optimal actuators for Optimus years ago. Tesla already produces its own sensors, its own batteries, its own chips (with Dojo). While OpenAI is recruiting its first hardware engineers, Tesla already has generations of prototypes behind it.
Figure AI : the former ally turned competitor
The history between OpenAI and Figure AI is a key chapter. OpenAI invested in Figure AI and provided the initial brain for the Figure 01 robot. But in February 2025, Figure broke off the collaboration to develop its AI in-house, notably with the Helix system that Figure AI launched to bring humanoid robots into homes.
This breakup is crucial. It means that OpenAI is entering robotics without any established hardware partner. Figure AI took the training data, the integration experience, and is heading towards its own model. OpenAI has to rebuild everything from scratch on the hardware side.
Genesis AI : the outsider pushing innovation
Genesis AI and its GENE-26.5 with humanoid robotic hands represent another approach: specialization on critical subsystems. Fine grasping, finger dexterity, is a problem that neither Tesla nor OpenAI have yet solved satisfactorily. Genesis is advancing on this specific front.
Comparative table of strengths
| Player | Main advantage | Weakness | Development stage |
|---|---|---|---|
| Tesla Optimus | Complete vertical integration, operational factories | AI software less advanced than OpenAI | Factory deployment |
| OpenAI Robotics | Best AI brain (GPT-5.5), world models | Zero hardware experience, no partner | Recruitment |
| Figure AI (Helix) | Robotics experience, first to market | Limited resources vs Tesla/OpenAI | Consumer launch |
| Genesis AI (GENE-26.5) | Hand dexterity, mechanical innovation | Limited ecosystem | Advanced R&D |
The frontier of embodied intelligence: why it's different from software
Embodied AI is not simply AI in a body. It is a fundamentally different paradigm. A language model can get a historical fact wrong without consequence. A robot that misinterprets the distance between its hand and a face causes an accident.
The first difference is real time. An LLM like GPT-5.4 Pro can take several seconds to generate a complex response. A falling robot must react in milliseconds. The very architecture of reasoning must change: shifting from sequential to parallel, from "think then act" to "perceive, act, reflect simultaneously".
The second difference is data scarcity. LLMs have been trained on the entirety of the internet's text. Robotic data — real physical interactions, grasping failures, falls — are extremely rare and expensive to produce. This is where OpenAI's world models become strategic: they make it possible to generate synthetic training data for the robot, a sort of "internet text" but for physics.
The third difference is safety. Analytics India reports that OpenAI is looking for ops and systems engineers, which is not trivial. Robotic systems require formal safety guarantees that LLMs have never had. A hallucination in a chatbot is embarrassing. A hallucination in a 70 kg robot is a legal and human risk.
The recruiting strategy: 10K salaries and Altman's message
Sam Altman's message on May 31, 2026, is remarkable for its simplicity. No press release, no detailed blog post. An X post, a single vision statement ("imagine everyone having a personal robot"), and a link to the job listings.
The advertised salaries for senior engineers go up to $10,000 per month according to feedback from the tech community. It's aggressive but necessary. OpenAI isn't just competing with the salaries of Google DeepMind or Anthropic. It's competing with Tesla, Boston Dynamics, Figure AI, and the entire San Francisco Bay Area robotics ecosystem.
The "full-stack hardware, ops, systems, and ML engineers" profile as reported by Clanks is unusually broad. Generally, a hardware engineer doesn't do ops, and an ML engineer doesn't do embedded systems. OpenAI is looking for hybrid profiles, capable of understanding the entire chain. It's rare and expensive.
Storyboard18 frames this announcement as a signal of expansion beyond software AI, a direct confrontation with Musk. The irony is that Musk co-founded OpenAI before being ousted, and he is building exactly what OpenAI is now starting to build, but with a several-year head start.
World models and robotics: the logical bridge between simulating and acting
OpenAI's world simulation program is the cornerstone of this strategy. To understand why, one must grasp what a world model is in the context of 2026.
A world model learns the laws of physics from data. It knows that a dropped glass falls, that a door has hinges, that a carpet provides friction. Fed with generations of images and videos, these models develop an implicit understanding of 3D space, gravity, and materials.
When you have a sufficiently accurate world model, you can use it in two ways. The first: generate simulated environments to train a robot without putting it in physical danger. The second: use the world model as a real-time prediction engine while the robot acts — "if I push this door with this much force, what will happen?"
Singularity Moments positions 2026 as the year when humanoids transform manufacturing. OpenAI is arriving exactly at this tipping point, with a theoretical advantage on the software front but a considerable delay on the hardware front.
The real question is whether OpenAI's world model is accurate enough for the robotics domain. Simulating a falling glass in a video is one thing. Predicting the exact force required to grasp a fragile object without breaking it, factoring in the tolerances of a real actuator, is another. The gap between simulation and reality (sim-to-real gap) remains the central problem of modern robotics.
The impact on the AI ecosystem: when software eats hardware
OpenAI's entry into robotics is redefining industry alliances. Until now, the pattern was clear: robotics companies (Figure, 1X, Agility) bought OpenAI's software AI and focused on hardware. That model is dead.
Figure AI was the first to realize this by ending its partnership in February 2025. Other robotics companies must now ask themselves an uncomfortable question: Will OpenAI become their customer, their competitor, or crush them?
For competing model providers, this is an opportunity. Anthropic with Claude Opus 4.7 (Adaptive) could become the brain supplier for robotics companies that refuse to depend on OpenAI. Google with Gemini 3 Pro Deep Think already has strengths in spatial understanding thanks to its experience with Google DeepMind and multimodal models.
The historical parallel is illuminating. When Apple started designing its own chips, processor suppliers had to reposition themselves. When Tesla started building its batteries, cell suppliers had to adapt. OpenAI in robotics is the same movement: the customer becomes a vertical competitor.
Moneycontrol summarizes the situation: Altman is about to face off directly with Tesla Optimus. But it's an asymmetric fight. Tesla has the factory, OpenAI has the brain. The question is which advantage is harder to catch up on.
What if OpenAI's robot also managed your finances?
The convergence between personal robotics and general-purpose AI assistants is inevitable. A personal robot living in your home will not just be a mechanical arm. It will be a mobile cognitive agent, capable of seeing your environment, understanding your habits, and acting on your daily life.
In this context, the trend toward financial agentic AI takes on a new dimension. ChatGPT becomes your banker with personal finance management for Pro users shows that OpenAI is already pushing its models into highly personal domains. A robot that knows your physical intimacy and your financial data creates a level of dependency without precedent.
It's also a commercial argument. If your personal robot can both tidy up your kitchen, monitor your spending, and plan your meals based on your budget, the perceived value explodes. OpenAI would no longer be selling a robot. It would be selling a complete life assistant, subscribed to monthly, with a business model similar to that of its current Pro offerings.
❌ Common mistakes
Mistake 1: Thinking OpenAI is starting from scratch in robotics
OpenAI had a robotics division between 2016 and 2021, before closing it to focus on LLMs. The team notably developed a robotic arm capable of solving a Rubik's Cube with an anthropomorphic hand. The institutional experience exists, even if the team from back then has left. It's not an absolute beginning, but it's almost a restart.
Mistake 2: Confusing LLMs and motor control
A model like GPT-5.5 is excellent at planning a sequence of actions. But generating text at 50 tokens per second has nothing to do with driving 40 actuators at 1000 Hz. The planning brain and the motor controller are two fundamentally different problems. OpenAI knows this, but the public often forgets it.
Mistake 3: Underestimating the gap behind Tesla
Tesla Optimus is already walking in Tesla factories. The company produces its own actuators, its own sensors, and has an assembly line ready to produce thousands of units. OpenAI is hiring its first hardware engineers in June 2026. Even with unlimited resources, catching up on 5+ years of vertical integration will take time.
Mistake 4: Believing that the world model solves everything
The sim-to-real gap is the nightmare of roboticists. A simulated world is always cleaner, more predictable, and smoother than the real world. Textures, frictions, and mechanical inaccuracies create gaps that the model cannot anticipate. World models reduce the problem, they don't erase it.
❓ Frequently asked questions
Who leads OpenAI's robotics division?
Aditya Ramesh, co-creator of DALL-E, leads the division. This is a strategic choice, signifying that visual and spatial understanding is considered the bridge between generative AI and robotics.
Which AI model will serve as the robot's brain?
There has been no official announcement, but GPT-5.5 (the first agentic LLM at 98.2) is the most likely candidate for high-level planning, coupled with faster specialized models for real-time motor control.
Has OpenAI done robotics before?
Yes, between 2016 and 2021, before shutting down the division to focus on GPT-3 and DALL-E. The institutional experience exists, but the team and technology have evolved since then.
What is the salary offered by OpenAI for these positions?
Listings show salaries of up to $10,000 per month for senior engineers, an aggressive level to compete with Tesla, Google DeepMind, and Bay Area robotics startups.
Why did Figure AI break ties with OpenAI?
In February 2025, Figure AI ended its partnership to develop its robotics AI in-house, notably with Helix. This breakup leaves OpenAI without a clear hardware ally at the time of launching its robotics division.
When will we see a commercial OpenAI robot?
No date has been announced. The current stage is purely recruitment. Given the lag behind Tesla and the lack of a production line, a functional consumer prototype is likely at least 3-5 years away.
✅ Conclusion
OpenAI is not getting into robotics to diversify. It is getting into it because software AGI is reaching its limits without a body to act in the real world. With Aditya Ramesh at the helm, world models as the foundation, and GPT-5.5 as the potential brain, the logic is coherent. But the lag behind Tesla Optimus is real, the lack of a hardware partner is a handicap, and the sim-to-real gap remains a scientific wall. The race for embodied intelligence is officially on — it will be as long as the race for the automobile.