Kael Zhang
OpenAIRoboticsEmbodied IntelligenceSam AltmanAGIPhysical AIIPO

OpenAI Reboots Robotics After Six Years: From Digital Intelligence to the Physical World

Kael Zhang

June 1, Sam Altman posted a job listing on X.

No press conference, no press release—just a tweet: OpenAI Robotics is hiring, looking for hardware, operations, systems, and machine learning engineers.

This tweet marked OpenAI’s official return to the physical world, six years after dissolving its robotics team in 2020.

Altman put it directly:

“Artificial intelligence should be able to help people in the real world.”

Short-term goal: build a batch of robots to assist technical workers in constructing infrastructure. Long-term vision: everyone owns a robot of their own.

Why 2026?

OpenAI is not new to robotics. Between 2016 and 2019, it launched the OpenAI Gym and Roboschool simulation platforms, and developed Dactyl, a dexterous robotic hand that could solve a Rubik’s cube with one hand.

In 2020, the team was disbanded. Co-founder Wojciech Zaremba gave a pragmatic reason: robotics training data was scarce, iteration was slow; whereas text and image data on the internet was abundant and easy to obtain. The decision to go all-in on large language models ultimately led to ChatGPT.

Six years later, the situation has completely reversed.

Large models are strong enough: GPT-4o and GPT-5 possess visual understanding, voice interaction, and code generation capabilities, providing a unified perception-understanding-planning foundation for robots.

VLA paradigm validated: Models like Google’s RT-2 and Physical Intelligence’s π0 prove that large models can serve as a robot’s “general policy network,” achieving cross-task generalization.

Industry ecosystem mature: Companies like Figure AI, 1X Technologies, and Unitree have validated the commercial viability of humanoid robots; capital and markets are in place.

IPO narrative needs it: OpenAI plans to go public in September 2026, valued at $852 billion. The story of “doing AI only in the digital world” lacks sufficient imagination. The physical world is a collection of trillion-dollar markets—manufacturing, logistics, healthcare, home services.

From Investment to In-House: The Breakdown of the Figure AI Partnership

OpenAI did not completely abandon the robotics track. After disbanding its in-house team, it invested in multiple robotics companies through its startup fund, including 1X Technologies, Figure AI, and Physical Intelligence.

The partnership between OpenAI and Figure AI, announced in February 2024, drew the most attention. OpenAI not only participated in Figure AI’s $675 million Series B funding but also developed a dedicated multimodal AI model for its humanoid robot. Just 13 days after the partnership, Figure 01 equipped with OpenAI technology demonstrated natural language interaction and autonomous operation capabilities.

But the partnership fell apart within a year. In February 2025, Figure AI founder Brett Adcock announced the termination, citing that “general-purpose large models cannot adapt to the hardware needs of robots; vertically integrated end-to-end models must be built.”

This event directly pushed OpenAI to “resurrect” its internal robotics team. From “behind-the-scenes support” to “front-stage player,” from “investment” to “strategic business.”

Team Configuration: Not Research, But Product

From the job openings at OpenAI Robotics, this is not a tentative “exploration project.”

Open positions include:

Job descriptions cover the complete hardware flow from concept exploration, prototype design, circuits, and PCB to integrated deployment, as well as simulation systems involving toolchains such as PhysX, MuJoCo, Unity, Unreal, and Omniverse.

Leading the team is Aditya Ramesh—co-creator of DALL·E 2 and one of the main developers of Sora. The team originated from an internal “world simulation” (Worldsim) research project, with the core idea of having AI first understand the operating laws of the physical world before putting it into a real robot body.

This is a typical “from zero to one” team configuration, meaning OpenAI is building a product, not a paper.

Technical Approach: Brain First, Body Later

OpenAI’s robotics strategy differs from the industry’s mainstream path.

Traditional robotics companies start with the hardware body and then overlay software capabilities. OpenAI chooses “brain first, body later”: first training AI through simulation environments to master the logic of physical world operation, then injecting that capability into physical robots.

This software-defined-hardware model, if successful, could disrupt the industry’s R&D paradigm. But the challenges are equally severe:

Figure AI’s Brett Adcock parted ways with OpenAI precisely based on these concerns. OpenAI now needs to prove that the general-purpose large model route is equally viable in robotics.

Competitive Landscape: The Warring States Era of Physical AI

OpenAI is entering an already crowded track.

PlayerApproachProgress
TeslaOptimus humanoid robot, self-developed end-to-endDeployed in factories, planned production of 50,000-100,000 units in 2026
Figure AISelf-developed end-to-end model, partnered with BMWFactory sorting tasks, 200 hours of continuous operation
1X TechnologiesHome-scenario humanoid robotEntered home testing phase
NVIDIAIsaac GR00T open platformProvides software stack and reference robots, does not build hardware itself
UnitreeCost-effective humanoid robotSTAR Market IPO imminent, shipments exceeding 10,000 units
Zhiyuan RoboticsDomestic embodied intelligenceCommercial deployment accelerating

OpenAI’s advantage lies in the “brain”—it currently possesses one of the strongest multimodal large models. Its disadvantage lies in the “body”—hardware engineering, supply chain management, and manufacturing capabilities are starting almost from scratch.

The essence of this competition is a debate between the “software-first” and “hardware-first” routes. OpenAI has a crushing advantage on the software side, but robotics is a domain where software and hardware are deeply coupled, and a pure software mindset may hit walls.

Strategic Move Before IPO

The timing of OpenAI’s robotics reboot is meaningful.

The company secretly submitted its S-1 prospectus draft to the SEC on May 22, planning to go public as early as September 2026. Deutsche Bank predicts the IPO valuation could exceed $1 trillion. But OpenAI faces severe financial realities:

The robotics business has been given a new mission: expand commercial boundaries through the integration of software and hardware, and tell the growth story of “from the digital world to the physical world” to the capital markets. The physical world’s market size is dozens of times that of the digital world; this narrative is crucial for supporting a trillion-dollar valuation.

Core Judgments

The establishment of OpenAI Robotics releases three key signals:

  1. Large model capabilities are spilling beyond the screen: GPT series capabilities in the digital world are approaching a ceiling; the physical world is the next battlefield
  2. The IPO narrative needs a new story: Relying solely on API and subscription revenue to support a trillion-dollar valuation leaves insufficient room for growth imagination
  3. The battle between general vs. vertical routes: Figure AI’s “defection” and OpenAI’s “return” represent the game between two technical approaches

Altman’s vision of “one per person” sounds distant, but OpenAI has no shortage of funds and talent to turn distant visions into reality. When it abandoned robotics in 2020, no one predicted that six years later it would return to the track with an $852 billion valuation.

This time, it brings not just money and models, but also the certainty of the entire AI industry’s migration from digital intelligence to physical intelligence.

Sources