[Series1]OpenAI Disbands Robotics Team as Physical AI Race Shifts
OpenAI folds robotics into its general-purpose AI as NVIDIA, Google, Figure and Tesla build their own ecosystems
Science Monitor Special Series | The Global Race for Physical AI, Part 1
The center of gravity in the global artificial intelligence (AI) industry is shifting from the digital world to the physical one. As general-purpose AI models that once handled language, images and code expand into physical AI, taking on visual perception, spatial reasoning and real-world action for robots, the way the robotics industry develops technology is also starting to change.
OpenAI’s decision to disband its standalone robotics team and fold its staff into core research and applied organizations is a symbolic sign of this industry shift. It is being read as a strategic turn away from building separate AI models for robots and toward applying the intelligence of general-purpose foundation models directly to physical robots.
Yet global rivals have not all chosen the same path. NVIDIA is building a platform that integrates robot learning, simulation and computing infrastructure, while Google DeepMind is advancing robot-specific VLA (Vision-Language-Action) models. Figure AI is feeding data gathered on real factory floors into its own models, and Tesla announced that it began producing its Optimus humanoid robot in October 2026.
In this seven-part special series, Science Monitor examines, step by step, the technology race, commercialization structures, data acquisition, AI chips and computing infrastructure, and national industrial strategies shaping the global physical AI industry. This first installment starts from OpenAI’s reorganization and Tesla’s move into production to look at a new industry landscape in which general-purpose AI models and robot-specific technologies compete.
OpenAI merges robotics research into its general-purpose AI organization
OpenAI recently disbanded its standalone robotics team and reassigned its researchers to other parts of the company.
According to a Nov. 13 (local time) report by the U.S. business magazine Fortune, the robotics research team moved into the core research organization led by Chief Research Officer (CRO) Mark Chen. The applied robotics team and the hardware team were placed under the Applied organization overseen by President Greg Brockman.
The company said the move is not a retreat from robotics. It cited the ability to apply general-purpose models, including GPT-6 Astra, directly to robotics as the reason for the reorganization.
OpenAI’s robotics organization has been wound down once before. The company shut down its early robotics team in 2020. In 2024 it rebuilt a standalone robotics team and agreed to co-develop AI models for robots with humanoid startup Figure AI. That partnership ended in February 2025, when Figure said it would develop its own robot AI.
The figures cited as evidence come from simulation
According to the company, NVIDIA research found that an agent based on GPT-6 Astra achieved a 74.8% success rate on a simulated robot manipulation benchmark. The best score by a robot-specific model on the same leaderboard was 51.0%.
These figures come from a simulated environment rather than real robots, and expert views are mixed. Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, told Fortune the reorganization signals that OpenAI sees robotics as an application area of general intelligence that is “no longer a special problem.” MIT professor Russ Tedrake said robots deployed in the field collect data from those sites, and that this data becomes more valuable as foundation models advance. Tedrake is co-founder and CEO of robotics company Walden Robotics [check spelling].
NVIDIA moves first on infrastructure for robot AI development
If OpenAI is focused on applying general-purpose AI models to robots, NVIDIA is building a platform that connects every stage of robotics, from learning and simulation to inference and deployment.
NVIDIA announced an expansion of its Isaac robotics platform at its GTC event in Berlin, Germany, on Oct. 21. The update includes Isaac OS, an operating system for robots; Isaac Lab 3.0, a robot learning framework; a reference platform for humanoid development; and data and simulation workflows.
Particularly notable is a development workflow that combines ChatGPT Astra with NVIDIA Isaac and Cosmos to accelerate robot data generation, training and evaluation. It shows that general-purpose AI models and robot-specific training platforms are not necessarily only competitors.
NVIDIA’s Isaac GR00T provides foundation models and development tools for humanoid robot reasoning and action. It is designed to take a robot’s camera images, natural-language commands and current state as input, generate continuous actions, and evaluate them on real robots and in simulation. Onboard computing platforms such as Jetson Thor provide the basis for running AI inference and control inside the robot.
The industrial significance of this strategy is considerable. It allows NVIDIA to supply computing and software to a wide range of robotics companies without depending directly on the success or failure of any single humanoid maker. It can be seen as a strategy to go beyond supplying GPUs and secure the standard platform for the robot AI development ecosystem.
Google DeepMind combines general reasoning with robot action
Google DeepMind is stepping up development of VLA (Vision-Language-Action) models that translate the intelligence of general-purpose models into robot actions.
Google DeepMind unveiled Gemini Robotics 2.5 on Nov. 10. The company introduced it as its most advanced VLA model, saying it improves the range of tasks robots can learn, learning speed, manipulation skills, and reliability and safety.
A VLA model goes beyond understanding visual information and natural-language commands to generate actual actions. For example, it is used when a robot recognizes an object on a workbench, interprets a user’s command, and then picks up and moves the object.
The difference between OpenAI and Google DeepMind lies in how they relate general-purpose models to robot-specific training. OpenAI emphasizes applying the capabilities of general-purpose frontier models directly to robots. Google DeepMind, alongside advancing its Gemini family of models, is separately developing a VLA framework for learning and evaluating real robot actions.
Both companies share the goal of extending general-purpose AI capabilities into the physical world, but they take different technical approaches to how robots generate actions and adapt to real environments.
DeepMind also ran a proof of concept (PoC) with robotics company Apptronik and Mercedes-Benz. Apptronik’s Apollo 2 humanoid was deployed on operating car production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The company said it achieved a 96% success rate in a task of sorting parts from bulk bins onto trays in each vehicle’s sequence.
The trial used Gemini Robotics ER, a reasoning model that tracks the workflow, together with Gemini Robotics 2.5, which controls hand and arm movements. Unlike OpenAI’s description of relying on a single general-purpose model, this structure splits reasoning from action control.
Figure AI and Tesla build up data in real-world settings
Unlike general-purpose AI model developers, humanoid makers are pursuing a strategy of accumulating data by combining hardware with real work environments.
After its partnership with OpenAI ended, Figure AI built its own technology ecosystem around its in-house robot AI model, Helix. It is pushing industrial deployment by pairing robot hardware, including its Figure 03 humanoid, with its own AI models.
Deployment on real factory floors is an important yardstick for robot AI competitiveness. BMW has disclosed its experience using Figure’s Figure 02 humanoid at its plant in Spartanburg, South Carolina. According to BMW, the robot moved more than 90,000 parts and operated for about 1,250 hours over 10 months. This is an example showing that humanoids can move beyond the lab into repetitive manufacturing work.
Tesla starts early production of Optimus
In its third-quarter shareholder update released on Oct. 21, Tesla said it had begun producing Optimus on a production line at its factory in Fremont, California. TechCrunch reported that the space was previously used for electric vehicle production.
Tesla described the robots coming off this line as its first “production-quality” units. It also said it had demonstrated that its current manufacturing process is suitable for a commercial product. It plans to keep refining the hardware and the process.
Production volume, pricing and the generation of the model being produced were not disclosed. The factory status table in the shareholder update listed California as “Early Ramp” and Texas as “Construction.” Neither site had a production capacity figure.
CEO Elon Musk called the current stage “initial production” on the earnings call. He said, “Since it’s an entirely new product, the initial ramp will be gradual and somewhat bumpy.” He put the timing for supplying outside companies at around the end of next year. He also said there is still work to be done on both the design and the AI.
Can general-purpose AI replace robot-specific models?
The most important question raised by OpenAI’s reorganization is whether general-purpose AI models can replace robot-specific models.
OpenAI chose to apply general-purpose models directly to robots. Google DeepMind splits reasoning and action into separate models and uses them together. NVIDIA supplies training, simulation and onboard computing as a platform. Figure and Tesla have integrated models, hardware and production in-house.
General-purpose models are strong at language understanding, complex reasoning, visual perception and task planning. Real robots, however, must deal with physical constraints such as millisecond-level response, precise joint control, balance and collision avoidance. Applying large cloud models to every step of control can be inefficient in terms of latency, network dependence and power consumption.
General-purpose models and robot-specific AI: hierarchical integration rather than competition
As physical AI advances, the roles of general-purpose large language models (LLMs), VLAs, world models and real-time control systems are being reshaped.
LLMs are strong at understanding natural-language commands and planning complex tasks. VLMs combine camera images and language to understand the surrounding environment. VLAs turn this information into actual robot actions.
World models are used to predict changes in the physical environment and the outcomes of actions. They predict what will happen when a robot pushes or lifts an object and assess risks before the action is carried out.
These models, however, do not handle everything down to real-time control of motors and joints. Balance, collision avoidance and precise position control require separate control algorithms and sensor-processing systems.
Physical AI may therefore evolve into a layered structure that combines general-purpose and robot-specific models. A general-purpose model in the cloud would understand complex commands and plan tasks, while lightweight models and control systems inside the robot would handle visual perception, action decisions and real-time movement.
In this process, on-device AI and edge computing become more important, because robots must be able to make a certain level of decisions and exercise control even without a network connection. Low-power NPUs, memory bandwidth, real-time sensor processing and functional safety technologies could become key factors determining the commercialization of humanoid and industrial robots.
Ultimately, rather than a structure in which a single general-purpose model controls every robot, the physical AI market is likely to develop into an ecosystem that combines general-purpose foundation models, robot-specific VLAs, simulation platforms, on-device AI chips and real industrial data.
For now, it is too early to judge whether the general-purpose or the specialized approach has the upper hand. The yardsticks for comparison going forward are likely to be task success rates, uptime and cost figures in real-world settings.
Series Preview | The Global Race for Physical AI
Part 1. OpenAI disbands its robotics team: Is the axis of physical AI competition shifting?
Part 2. NVIDIA Isaac and GR00T: The race for robot AI platform dominance
Part 3. The humanoid data war: Real industrial sites are training AI
Part 4. Tesla, Figure AI and Unitree: The race to mass-produce humanoids
Part 5. Robot NPUs and edge AI: The rise of on-device computing
Part 6. Manufacturing, logistics and defense: The economics of commercializing physical AI
Part 7. The U.S.-China race for physical AI and the reshaping of global supply chains
Related articles
- Google DeepMind unveils Gemini Robotics 2.5: Up to 100% higher success rates on some 200 tasks, 99.5% resistance to robot jailbreaks (in Korean)
- AI’s next challenge: Beyond language into the real world (in Korean)
- GPT-6 Astra proves 10 hard problems, including the Navier-Stokes problem, with AI (in Korean)









![[기획2]엔비디아 Isaac·GR00T, 로봇 AI 플랫폼 패권 경쟁](https://scimonitors.com/wp-content/uploads/2026/11/jetson-thor-1-1536x864-png-150x150.webp)