OpenAI Returns to the Robotics Race: A Late Gamble or a Long-Planned Bombshell?

OpenAI is hiring again — this time for humanoid robots.

In June 2026, multiple sources including QbitAI confirmed that OpenAI is recruiting four core positions for its mysterious robotics team: robotics control algorithm engineer, robotics hardware designer, embodied intelligence researcher, and simulation environment engineer. The news sent shockwaves through the industry — the OpenAI that quietly disbanded its robotics team in 2020, then quietly invested in competitors, has finally decided to get in the game itself.

Interestingly, just a few months ago, reports surfaced that OpenAI was developing a humanoid robot prototype internally, codenamed "Project Striker." The large-scale hiring now means this project has been upgraded from "testing the waters" to "dead serious." Why is a company that dominates the software world suddenly interested in hardware? Is the combination of large models and robots a real inflection point, or just a capital story?

I dug through the public information and found several key signals.

1. Background and Timeline

OpenAI's robotics ambitions go back further than most people think.

In 2019, OpenAI briefly explored robotics research, attempting to train robotic arms to complete simple tasks using reinforcement learning. But at the time, large model capabilities were far from sufficient, and the cost of data collection was prohibitively high. The project was strategically shelved in 2020. Core team members either transferred internally or left to join competitors.

The turning point came in February 2024. OpenAI suddenly announced a partnership with Figure AI, unveiling the industry-shocking Figure 01 — a humanoid robot that could hold conversations, understand intent, and complete complex household tasks. In the video, a human says "I'm hungry," the robot glances at an apple on the table, and hands it over directly. The whole process takes less than 10 seconds. The demonstration went viral on social media, making everyone realize for the first time: large models can truly make robots "understand" humans.

What followed was even more interesting. OpenAI publicly stated "we don't make hardware" while quietly investing in Figure AI (investment amount reportedly approximately $675 million, according to TechCrunch), 1X Technologies ($125 million), and other robotics companies. This isn't "not making hardware" — this is "letting others pave the way first, then stepping in when the timing is right."

The June 2026 large-scale hiring is that "timing is right" signal. The four positions cover the full stack from low-level control to high-level intelligence: control algorithms determine how the robot moves, hardware design determines what the robot looks like, embodied intelligence research determines how the robot "thinks," and simulation environments are used for training and testing — without simulation environments, the cost of real-world training would be prohibitively expensive.

OpenAI is serious this time.

2. Core Technical Principles

Large models plus robots sounds impressive, but the principles aren't that mysterious.

Traditional robots rely on "preset rules": engineers write code telling the robot "if you encounter obstacle A, go around to the left." Encounter a scene it hasn't seen before? Sorry, it doesn't know what to do. What OpenAI wants to do is make robots "understand" tasks like humans, then figure out how to complete them on their own.

There are two key technologies behind this.

The first is Vision-Language-Action models (VLA). Think of it as a multimodal upgrade of GPT — it doesn't just look at images and describe them, it can also "output actions." Figure 01 uses this architecture: cameras capture images, the large model understands "what is this and what should be done," then generates specific action commands for the robot to execute. With this entire chain connected, robots can truly "follow instructions."

The second is simulation training plus real-world fine-tuning. Training robots in the real world is extremely expensive — breaking a single component can cost thousands of dollars, and training a somersault skill might require tens of thousands of falls. So the industry-standard approach is to train in simulation first, then validate on real hardware once the performance is good enough. OpenAI's hiring of simulation environment engineers means they're building their own "digital twin training ground."

Key technology checklist:

• VLA model architecture: Unifying visual understanding, language reasoning, and action generation in a single model — this is the core breakthrough in embodied intelligence.
• Few-shot generalization: Large models enable robots to handle unseen objects in a "zero-shot" manner, without programming each task individually.
• Real-time inference latency: Current VLA models have an end-to-end response time of approximately 100-200 milliseconds (according to Figure AI's public technical blog), with room for improvement.
• Sim2Real transfer: How to effectively transfer skills learned in simulation to real robots — this is a widely recognized technical challenge in the industry.
• Tactile sensing fusion: Beyond vision, tactile feedback is critical for fine manipulation. Current mainstream solutions include novel tactile sensors like GelSight.

One thing must be made clear: the technical principles are easy to explain, but engineering implementation is another matter entirely. OpenAI's advantage lies in its large model capabilities, but in areas like hardware control and mechanical design, they are true newcomers. So this hiring round will likely poach talent from teams like Boston Dynamics and Tesla Optimus.

3. Why This Matters

OpenAI making robots isn't just one company's strategic choice — it's an event that could reshape the entire industry landscape.

why it matters for OpenAI itself. ChatGPT's success proved the value of large models, but the ceiling is also obvious — it's just a conversational tool with no way to directly impact the physical world. If OpenAI wants to continue growing its valuation and expanding its influence, it must find new outlets for large models. Robots are a natural carrier: they turn AI from "all talk" to "action," from the cloud to the real world.

More importantly, robots may be a necessary path to AGI (Artificial General Intelligence). OpenAI's ultimate goal is AGI, and true general intelligence must be able to understand the physical world and complete physical tasks. OpenAI CEO Sam Altman has stated publicly on multiple occasions that embodied intelligence is one of the key paths to AGI. This round of hiring and expansion isn't a whim — it's strategic positioning.

For the industry, what does OpenAI's entry mean? It means more capital, stronger technology, and fiercer competition. Will companies like Figure AI and 1X Technologies, which already received OpenAI investment, become collateral damage? Will Tesla's Optimus and Figure AI's own products face greater pressure because OpenAI is now competing directly?

There's another overlooked point: data. Large models need massive amounts of training data, and high-quality data in the robotics field is extremely scarce. OpenAI has internet-scale data, but robot operation data is virtually nonexistent. If they can build a sufficiently large robot data flywheel — like they did with text data — latecomers will struggle to catch up.

This isn't an ordinary competition. This is a war over "who defines the next generation of human-computer interaction."

4. Industry Impact and Data

The embodied intelligence sector is experiencing an unprecedented capital boom.

According to a 2025 report by Markets and Markets, the global embodied AI market is expected to grow from approximately $3.1 billion in 2024 to $13.8 billion in 2028, with a compound annual growth rate exceeding 45%. Goldman Sachs is more optimistic, predicting that by 2035, the humanoid robot market alone could reach $38 billion.

Capital has the sharpest nose. In 2024, global investment in the robotics sector was approximately $6.7 billion according to Crunchbase, a 32% year-over-year increase. The proportion of funding going to embodied intelligence-related companies surged from 8% in 2022 to 35% in 2024. Figure AI alone raised over $700 million in 2024, with its valuation reaching $2.6 billion at its peak.

Key data points:

• Figure 01's demonstration video received over 5 million views within 24 hours of release (according to Figure AI's official Twitter), becoming one of the hottest tech topics of the year.
• 1X Technologies' NEO Gamma robot underwent over 2,000 hours of home environment testing in Norway (according to 1X's official blog), making it one of the most aggressively deployed humanoid robots currently available.
• Tesla Optimus is expected to enter mass production in 2026. Musk claims the final cost will drop below $20,000, but analysts generally believe scaled delivery before 2027 is unlikely.
• Boston Dynamics' electric Atlas has entered the commercial pilot stage, but the high cost of its traditional hydraulic solution remains a barrier to scaling.

The industry landscape is also changing. Before 2022, robotics startups had difficulty raising funding, with capital favoring software and AI applications. But starting in 2024, the hardware-plus-AI combination suddenly became attractive. The reason is simple: large models can finally "read" the physical world. The robot's "brain" problem is solved. What's left is the body problem.

The body problem happens to be exactly what hardware companies are good at. So in the coming years, we'll see more and more "software company + hardware company" partnerships, or simply capital-led mergers and acquisitions. OpenAI's choice to build its own team rather than continue investing suggests they believe the time window has matured.

5. Real-World Case Studies

Case Study 1: Figure AI, From Near-Bankruptcy to $2.6 Billion Valuation

Figure AI was founded in 2022. Founder Brett Adcock is a serial entrepreneur who worked on electric aircraft and VTOL vehicles before going all-in on humanoid robots in 2022. The early days were brutal — investors thought it was too expensive and too sci-fi, and the team almost couldn't make payroll.

The turnaround came in 2023. Figure AI began reaching out to large model companies, attempting to integrate vision-language models into robot control systems. The Figure 01, launched in partnership with OpenAI in February 2024, made everyone remember the company overnight. Adcock later said in an interview that the demonstration video actually only showed 10% of the robot's capabilities, but that was enough to make investors go wild.

The numbers speak for themselves: Figure AI completed over $700 million in funding in 2024, with its valuation skyrocketing from $200 million to $2.6 billion. The company grew from 40 employees in 2023 to over 300 in 2024, and established a dedicated robotics training laboratory in California.

Currently, Figure 01 has partnered with BMW for pilot deployment on automotive manufacturing assembly lines, primarily handling parts transportation and material sorting tasks. According to BMW's official press release, the first pilot runs exceeded six months with a task success rate of approximately 87%.

Case Study 2: 1X Technologies, The Aggressive Play in Home Scenarios

Unlike Figure AI's focus on industrial scenarios, 1X Technologies targeted the home scenario from the start. Their flagship product, NEO Gamma, is a humanoid robot standing 1.65 meters tall and weighing 30 kilograms, positioned as a household assistant.

1X's strategy is interesting: deploy extensively in real home environments first, then iterate the product by collecting real user data. They signed testing agreements with hundreds of families in Norway, with robots running in real homes for over 8 hours per day. This "surround the cities from the countryside" approach has allowed 1X to accumulate a large amount of real-world scenario data that competitors simply cannot obtain.

According to 1X's official data (released in Q4 2024), NEO Gamma's task completion rate in home environments has improved from an initial 45% to 72%. User feedback shows the most popular features are "room tidying" and "elderly companionship," while the least popular is "dishwashing" — because the current robotic hand's grasping success rate in wet environments remains relatively low.

OpenAI led 1X's Series B funding round in 2024 ($125 million), which was interpreted by outsiders as OpenAI's early positioning in the home scenario. 1X CEO Bernt Bornich once stated: "OpenAI isn't just investing in us — they stationed a technical team in our office, working together to solve the core problems of embodied intelligence."

6. Competitive Landscape

The robotics track is already packed with players, each with different strategies and advantages. Here's a horizontal comparison across multiple dimensions:

Several interesting observations from the table:

Figure AI and OpenAI are tightly bound. If OpenAI enters the water directly, Figure AI's independence could be questioned. But thinking about it the other way, OpenAI might prefer to continue supporting Figure and focus on the AI brain rather than making hardware themselves. Tesla Optimus's advantage is manufacturing capability and cost control, but Musk's promises have a high "delay rate." 1X has the clearest scenario positioning, but its funding scale is relatively small, which may limit expansion speed.

My judgment is: in the next 3 years, this sector will experience a brutal shakeout. The survivors will be those with unique data barriers, mature hardware supply chains, or enough money to burn. OpenAI falls into the third category, but money alone isn't enough — they need to prove they can make "usable" robots, not just another research paper.

7. Technical Challenges and Limitations

After all the good news, it's time to pour some cold water.

Hardware bottlenecks are more serious than imagined. Large models handle "thinking," but robots need to "act." Current robotic arms still have low success rates when grasping irregular objects and soft items. According to data from multiple academic papers and industry reports, existing robotic hands have a success rate of approximately 60-70% on fine manipulation tasks like "picking up a grape from a bowl," far below the human level (95%+). This isn't an algorithm problem — it's a gap in materials science and precision manufacturing.

The Sim2Real gap still exists. Skills trained in simulation often don't transfer well to real robots. Changes in lighting, differences in ground friction, and variations in object weight and texture can all cause significant performance degradation. Figure AI engineers revealed in a blog post that it took them six months to get simulation-trained policies to an acceptable level on real hardware.

Data collection costs are extremely high. High-quality data in the robotics field is extremely scarce, unlike internet text that can be obtained at low cost at scale. Training a robot capable of handling 100 household tasks requires collecting at least millions of demonstration data points. According to industry estimates, the production cost of a high-quality robot operation dataset is approximately $1,000-$5,000 per minute.

Safety concerns cannot be ignored. Robots operating in close proximity to humans in home environments could cause serious consequences if a control system bug causes the robotic arm to go out of control. The industry currently lacks unified safety standards and certification systems, which is a potential barrier to large-scale deployment.

The commercialization path is unclear. Figure 01's BMW pilot sounds good, but there's a vast gap between "pilot" and "scale." Robots need maintenance, require accessories, and need custom development. When you add up these costs, the ROI (Return on Investment) may look very unattractive.

In one sentence: Between a technical demo and commercialization lies an entire Himalayan mountain range.

8. Who Should Care About This

If you fall into any of the following categories, this directly affects you.

Developers: If you're learning machine learning or reinforcement learning, embodied intelligence is the next big direction. Getting in now and learning VLA models, Sim2Real transfer, and robot simulation environment development are all valuable tech stacks. OpenAI's job requirements can serve as a learning roadmap.

Product Managers: Robot products are completely different from internet products — hardware has physical limits, user expectations are higher, and the tolerance for error is lower. If you have ideas for robot products, now is the best time because the industry hasn't solidified yet.

Investors: The embodied intelligence sector is already hot, but valuations are generally high. Does Figure AI's $2.6 billion valuation have a bubble? Can Tesla Optimus really achieve mass production? These questions require deep research. My advice is to focus on companies with real data accumulation, not just projects that tell good stories.

Industry Professionals: If you work in manufacturing, logistics, healthcare, or elderly care, humanoid robots may soon affect your work. Watch competitor dynamics and think about how to collaborate with robots, rather than waiting to be replaced.

Entrepreneurs: If you want to build a complete robot, the barrier to entry is already very high. But if you want to be the "pickaxe seller" in the robotics gold rush — data collection tools, simulation platforms, testing services, custom development — there are still plenty of opportunities.

9. Future Predictions

A few predictions from me — if I'm wrong, consider it throwing bricks to attract jade. If I'm right, come back and give me a like.

First, OpenAI won't build hardware themselves. They'll likely continue investing in or controlling one or two robotics companies, focusing on the AI brain and datasets themselves. This is the same logic as Microsoft making Copilot rather than making PCs — OpenAI's strength is software; hardware is a different world.

Second, 2026-2027 will see a wave of robotics company closures. The sector is too hot right now, valuations are inflated, and when the capital winter comes, the first to fall will be projects without commercialization capabilities. The survivors will be companies that can clearly articulate "who will buy, why they'll buy, and can they afford it."

Third, home scenarios are harder to crack than industrial scenarios. Many people think robots will replace factory workers first, but I think the opposite — factory environments are controllable, tasks are standardized, and ROI is easy to calculate. Home environments are complex, user expectations are high, and the cost of accidents is significant. Figure AI's decision to start with industrial scenarios is the right call.

Fourth, 2028 could be a critical inflection point. If embodied intelligence technology hasn't made substantial breakthroughs by then, this wave of enthusiasm may cool for a while. Goldman Sachs' predicted $38 billion market requires sustained technological progress to materialize — right now it's still just a prediction.

Fifth, OpenAI's entry will significantly raise the competitive bar. Not every robotics company has the ability to train its own large models. The future may see a landscape where "companies with strong AI capabilities control the core brain, while hardware companies become contract manufacturers." Is this good or bad for the industry? Too early to say.

10. Actionable Advice

At the end of the day, readers need to know "what should I do?"

If you want to enter the embodied intelligence space, start learning VLA models and simulation environment development now. The time window is still 1-2 years. If you're already in the robotics industry, think carefully about where your differentiation advantage lies — is it data, hardware, or scenario understanding? Don't try to do everything; focus on one point and execute it to the extreme.

If you just want to follow this sector, don't let media hype lead you around by the nose. Figure 01's demo is cool, but it still has a long way to go before it truly changes lives. Look more at actual deployment data and less at PowerPoint presentations.

OpenAI returning to the robotics race is a significant event worth watching, but it's still too early to conclude whether they'll succeed. What we can be certain of: large models plus robots has become an irreversible industry consensus. The rest depends on who can turn this vision into reality.