Tech • AI • Robotics • Game
OpenAI's First AI Robot Is Finally Happening (ChatGPT Gets a Body)
TL;DR
OpenAI has confirmed plans to build its own humanoid robot and other robotic form factors, targeting a 2027 demo as it shifts from supplying AI models to competing directly in robotics hardware.
KEY POINTS
OpenAI sets a 2027 humanoid deadline
Sam Altman has said OpenAI will “definitely” build a humanoid and expects a public demo in 2027. He also said the company is pursuing “both” humanoids and robots designed for data centers or other practical settings. The strategy reflects a view that the human world is built around human bodies, making humanoid form factors useful for tasks such as opening doors, typing, driving equipment, and cleaning.
Industrial work comes before home robots
The company does not appear to be prioritizing a household robot in the near term. Early deployments are expected in industrial roles, especially building and maintaining data centers and manufacturing equipment. Altman has still framed the long-term goal as a personal robot for everyday chores, but near-term demand may come from OpenAI’s own infrastructure expansion.
A revived robotics unit is built around AI generation models
OpenAI formally launched its robotics division on May 31 and built it from World Sim, an internal simulation effort. The group is led by Aditya Ramesh, known for work on DALL-E and Sora. The technical bet is that robot control can be handled by a vision-language-action model, which uses the same broad architecture as image and video generation systems but outputs physical movements instead of text or visuals.
Real-world training data is now a central focus
Data shortages helped end OpenAI’s earlier robotics effort in 2020, so the new push includes a large data-collection operation. In March 2026, the company leased 202,000 square feet at Portside Commerce Center in Richmond, California, with 14,000 amps of power and space for 76 robot workstations. Human operators use handheld rigs to guide Franka robot arms, generating paired records of what the robot saw and how it moved.
Simulation remains useful but risky
World Sim is meant to produce large volumes of cheap training data, but robotics still faces the classic sim-to-real gap. Systems that perform well in virtual settings often fail when exposed to real lighting, friction, clutter, and sensor noise. That gap was visible at Nvidia’s GTC Berlin, where four small Pollen Robotics MicroDucks performed basic motions but several toppled over during live commands.
Hardware may be the hardest part
OpenAI’s hiring shows it is building a full hardware program, including actuators, sensors, electronics, firmware, simulation pipelines, and prototyping labs. Actuators alone can account for 40% to 60% of a humanoid’s parts cost, and custom versions often take two to four years to reach production. A 2027 demo may be achievable, but mass manufacturing is a separate challenge, especially for a company that has not produced hardware at scale.
Leadership and ethics questions have surfaced
The hardware effort has also seen turbulence. Caitlin Kalinowski, a former Meta executive hired in late 2024 to lead robotics hardware, resigned in March 2026 after reportedly raising concerns tied to defense work and safeguards around surveillance and lethal autonomy. OpenAI has not named a replacement, leaving a notable gap at the top of a highly complex hardware program.
The move upends former partnerships
By entering robotics directly, OpenAI is now competing with companies that once relied on or were funded by it. Figure AI ended its OpenAI collaboration on February 4, 2025, arguing that robotics companies needed to own their own AI stack, and later deployed Figure 03 at BMW. 1X Technologies, backed by OpenAI in 2023, now sits in a more awkward position as it develops the home-focused Neo robot while SoftBank explores a controlling stake at a $6 billion valuation.
The market is already moving fast, especially in China
About 19,100 humanoids shipped in the first half of 2026, up 272% year over year, with more than 97% coming from Chinese companies. AGIBOT held 44% of the market and Unitree 31%, mostly in factories and logistics. In the West, Tesla’s Optimus Gen 3 is already doing basic manufacturing work, while Google DeepMind has launched Gemini Robotics 2, underscoring a broad industry push to unite AI “brains” with robot “bodies.”
Security is becoming a competitive argument
Researchers at Binghamton University showed how a robot using a language-model planner could be manipulated by a hidden trigger. In the demonstration, adding the letters CF to an otherwise normal command caused a knife-wielding robot to redirect toward a model human hand. The test did not involve an untouched commercial product, but it reinforced a growing view that outsourcing a robot’s AI core may create serious security and safety risks.
CONCLUSION
OpenAI’s robotics push marks a major strategic shift from AI supplier to full-stack competitor spanning chips, data centers, devices, and now robot bodies. A 2027 demo looks plausible, but turning that into safe, scalable hardware will determine whether the company can catch more established robotics players.
Ask a question