On October 17, 2026, when OpenAI CEO Sam Altman was asked on the Sources podcast whether the company would really build humanoid robots, he gave a two-word answer: "We will." — not "We are considering it," and not "It's one direction."
OpenAI had never publicly committed to entering the robot hardware field. After its collaboration with robotics company Figure AI fell apart, it even led outsiders to believe the company would continue focusing on a pure-software path. This explicit statement marks the first time in OpenAI's history that it has characterized humanoid robotics as an in-house R&D project rather than an investment or partnership target.
"Brain" Problem: The Underestimated Core Obstacle
Altman's logic is not hard to understand: the infrastructure of human society—doorknobs, keyboards, tools—is designed around the human body. On the podcast he said, "The world is largely designed for humans, and matching that form factor seems reasonable." This is a plain engineering argument, but Altman immediately shifted the focus to a harder problem.
He believes the robot's physical details are "far less important than figuring out the brain that makes the robot run." This points to the hardest bone in robotics for decades: the complete closed loop of perception-planning-execution.
Engadget's report gives a concrete example: to pick up a mug, a robot needs to recognize the cup, locate the handle, estimate grip force, and then apply just the right amount of force to lift it. Too much force and the cup breaks; too little and it falls to the floor; miss the handle and coffee spills. Every step requires close coordination among the perception system, reasoning system, and motion system—and the real world will not wait for you to recalibrate.
OpenAI has historical accumulation in this field, but also historical lessons. The 2019 Dactyl project trained a robotic hand to learn to manipulate real objects in a simulated environment and used the same method to solve a Rubik's cube; it was regarded as an early model of simulation-to-reality transfer. But the project ultimately did not move toward productization. The "sim-to-real gap"—the phenomenon in which motions learned by a model often fail under real physical constraints—has always been a limiting factor.
27 Job Openings Reveal Strategic Intent
Compared with Altman's public statement, the 27 robotics positions on OpenAI's careers page say more. These roles cover motion control, state estimation, actuator control, whole-robot behavior coordination, and machine learning infrastructure for physical AI systems—a complete vertical integration path from algorithm research to hardware manufacturing and supply chain.
According to AI Weekly, four of the 27 positions are specifically focused on actuator design, a capability that traditional outsourcing or model-licensing partnerships would not require the company to build itself. Actuators are the robot's "muscles," determining joint torque, precision, and energy efficiency. In-house actuator development means OpenAI does not plan to outsource the robot's core physical components.
This choice has its own internal logic. Robot training requires large amounts of real-world physical interaction data, and if the robot itself is third-party hardware, data collection, quality, and ownership become complicated. According to AI Weekly's analysis, owning its own robot is essentially building proprietary physical AI training data infrastructure—something no partnership agreement can replace.
From a Failed Partnership to In-House Development
This shift did not come out of nowhere. OpenAI had partnered with humanoid robotics company Figure AI and participated in its $675 million funding round. But according to multiple media reports, the partnership ultimately ended when Figure AI voluntarily terminated it. In February 2025, Figure CEO Brett Adcock announced that the company would shift to in-house AI system development and launched its own VLA (vision-language-action) model Helix, reasoning that "we can't outsource robot AI, just as we can't outsource hardware."
Figure AI's choice inadvertently validated a judgment: as AI model capabilities advance rapidly, relying on an external AI provider puts robotics companies in a passive position. The conclusion OpenAI drew from this partnership is clearly the same logic—but in the opposite direction: rather than providing brains to others, it is better to build a complete robot itself.
According to Humanoids Daily, OpenAI also explored the possibility of acquiring Norwegian robotics company 1X (a company in which OpenAI has invested), but ultimately chose the in-house R&D path.
Competitive Landscape: How Far Hardware Companies Have Run While Software Companies Enter
By the time OpenAI announced its entry, the competitive landscape of the humanoid robotics track was no longer what it used to be.
Tesla's Optimus project continues to advance, but according to reports from outlets including Electrek, the project faces mass-production delays in 2026, with supply chain and end-effector precision issues as the main obstacles. Musk's original target of one million units per year has encountered friction against engineering reality.
Meanwhile, XPeng announced in September 2026 that its IRON humanoid robot had rolled off the line—literally "walked off the production line itself"—with a mass-production target set for the end of 2026. According to Engadget, IRON has 76 degrees of freedom (21 in each hand), and XPeng built a dedicated production line with an automation rate exceeding 80% for it.
From this horizontal comparison, OpenAI's 2027 "cool demo" means that while XPeng is already in mass production and Tesla is solving engineering problems, OpenAI is still at the demonstration stage.
Target Market: A Two-Step Path from Construction Sites to Living Rooms
Altman has said on multiple occasions, "I think everyone should have a personal robot," but OpenAI's job descriptions reveal a more pragmatic short-term positioning: the recruiting copy explicitly states that the near-term goal is "general-purpose robots that help skilled workers build infrastructure," rather than household robots.
This ordering matters. Industrial scenarios have relatively higher tolerance for operational errors, and the human-robot interaction environment is relatively controllable; home scenarios mean countless unforeseen variables and privacy and safety frameworks that do not yet exist. Altman himself admits that it will take several more years for robots to "go out onto the streets."
But between these two steps, one uncertainty is deliberately omitted: what is the commercialization path from industrial demonstration to home scenarios? Pricing models, developer ecosystems, third-party integration—all these questions are absent from Altman's statement. OpenAI has not released any design drawings, disclosed any technical roadmap, or given any commercial price range.
Independent Judgment
What is most noteworthy about OpenAI's entry into humanoid robotics this time is not "when it will be built," but "why now, and why in-house."
From the collapse of the Figure AI partnership to building an actuator team on its own, this shift in route shows that OpenAI judges that AI model capabilities have reached a critical point, and the value of a physical carrier is beginning to exceed the marginal benefits that pure software can bring. In other words, the next stage of AI capability improvement requires not only more text and code but also real-world physical interaction data—and that can only be collected by one's own robots.
But this path has a practical difficulty: the physical world iterates far more slowly than software. A model can be trained and updated within weeks, while a batch of robots takes years from design to manufacturing to deployment. OpenAI's accustomed "fast iteration" culture will face a real test under hardware constraints.
The 2027 "cool demo"—to use Altman's own word—is just a demo. The real moment of judgment will be whether, after this demo, OpenAI can present a path that can be mass-produced, profitable, and stably operating in the physical world. Before that point, this announcement is more like an admission ticket than a report card.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接