AI News

OpenAI Dissolves Independent Robotics Team as General-Purpose Models Show Mixed Results in Embodied Intelligence Tests

East MoneyUpdated Nov 14 14:59

AI summary

OpenAI has confirmed that it has dissolved its independent robotics team, with robotics researchers merged into the core research team and the applied robotics and hardware teams moved into the applied division; the company says robotics remains part of its long-term vision. Multiple tests compiled by The Paper show that an agent solution based on GPT-6 Astra achieved a 74.8% success rate on NVIDIA's RoboLab-120 simulation benchmark, higher than the 51.0% of the specialized model Lakeside; in UBTECH's six real-robot and simulation tests, Astra's direct control failed to complete grasping a Rubik's cube, simulated basket carrying, and direct bowl stacking, but when coordinated with the in-house Thinker-VLA it completed lucky-bag grasping, obstacle avoidance, and bowl stacking; in the bowl-stacking task, Astra's token consumption and estimated API costs were reduced by 83.6% and 74.2%, respectively. NVIDIA's research also notes that the direct-control approach takes an average of about 18.56 seconds per inference, and the highest-success-rate approach costs about $36.40 in API fees per successful task; latency and cost remain constraints on practical application.

Why read this

AI score 60

Simulation and real-robot tests by multiple institutions show that the division of labor between general-purpose models and specialized embodied models is affecting task completion rates and invocation costs.

Topics

Tags

Body · AI translation

Main text

The progress of general-purpose large models in robotics is prompting the embodied intelligence industry to rethink the R&D path for robot "brains."

Recently, in conversations with several embodied intelligence practitioners, The Paper found that the release of GPT-6 Astra has become an important milestone, and the capabilities it demonstrates have led the industry to focus more on which robot tasks can be handed directly to general-purpose models and which still require specialized training.

On November 13 local time, according to foreign media reports, OpenAI confirmed that it has disbanded its standalone robotics team, with robotics researchers merged into the core research team and the applied robotics and hardware teams transferred to the applications department. The company said robotics remains an important part of its long-term vision and that the adjustment is intended to accelerate related R&D.

From simulation testing to real-machine practice, general-purpose models are taking on more robot tasks, but fine control, response speed and operating costs remain issues that must be resolved before they can be put into practical application.

From simulation to real machines, general-purpose models begin to take on robot tasks

NVIDIA's recently published Astranomy research shows that on the RoboLab-120 simulated manipulation benchmark, an agent-based approach built on GPT-6 Astra achieved a 74.8% success rate, higher than the 51.0% of the specialized model Lakeside listed in the study.

Of these, the 74.8% corresponds to the agent-based approach with persistent interaction context; the approach that directly generates actions achieved a 66.3% success rate. The tests mainly covered manipulation tasks such as grasping and placing; the simulation environment was paused during model inference, so its ability to respond in real time in a continuously changing environment has not yet been tested, nor did the tests cover complex contact tasks such as insertion and assembly.

NVIDIA test results

Similar progress has also appeared with Anthropic's new models.

Anthropic's released Claude Fable 5.2 and Mythos 5.2 scored 64.9% and, under the same test framework, GPT-6 Astra scored 29.8% in the RoboDojo Sim dual-arm robot simulation test run internally by the company. The company also demonstrated Fable 5.2 controlling a robot to unscrew a bottle cap, paint and perform other tasks. This test uses a different benchmark from the one used in the NVIDIA study, so the scores cannot be directly compared.

Domestic robot companies have also put GPT-6 Astra into actual testing.

On October 22, at the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Unitree Technology founder Wang Xingxing said that physical AI has already shown signs of being broken through, and that a breakthrough may emerge in the next one to three years. On site, he demonstrated robot tests conducted by Unitree using GPT-6, involving tasks such as painting and planning navigation.

UBTech, meanwhile, tested both the path of direct control by a general-purpose model and the path of coordination with an embodied model.

Previously, The Paper learned that UBTech conducted six groups of real-machine and simulation tests around GPT-6 Astra, covering real-machine grasping of a Rubik's cube, grasping a lucky bag, and autonomous obstacle avoidance, as well as basket carrying, direct control of bowl stacking and bowl stacking combined with a world model in a simulated environment.

According to results disclosed by the company, Astra failed to complete direct grasping of the Rubik's cube, simulated basket carrying and direct bowl stacking; after coordination with a specialized embodied model, the robot completed lucky bag grasping, obstacle-avoidance movement and bowl stacking.

The coordinated approach involved calling UBTech's self-developed Thinker-VLA to execute operations, as well as combining Thinker-WM to predict action outcomes and assist decision-making. In the bowl-stacking test, compared with the direct-control approach, the coordinated approach reduced Astra's token consumption and estimated API costs by 83.6% and 74.2%, respectively.

UBTech test of Astra direct control grasping a Rubik's cube

These practices present different ways for general-purpose models to enter robotics: they can directly generate actions, or they can be responsible for understanding tasks and formulating plans, then calling specialized models for execution. How models divide labor among themselves has already begun to affect task completion and operating overhead.

General-purpose models participate in decision-making and R&D, while specialized models continue to fill in execution capabilities

The fact that general-purpose models can take on more tasks does not mean that specialized training for robots has become useless. Google's recently updated robotics system is still improving practical operating capabilities through the coordination of embodied reasoning and action models.

On November 10, Google DeepMind released Gemini Robotics 2.5. This is a vision-language-action (VLA) model that can translate visual observations and language instructions into robot actions. According to the company, the model was evaluated on more than 200 tasks, covering capabilities such as tool use and deformable object manipulation, including using tongs to pick up and place food, tying a bow on a gift box and moving liquids.

Google also disclosed that in a proof of concept for an automotive production line conducted with Apptronik and Mercedes-Benz, the Apollo 2 humanoid robot performed parts picking and kitting operations with a success rate of 96%. Among them, the embodied reasoning model was responsible for tracking the workflow, while the Gemini Robotics 2.5 action model controlled the arms and hands to take parts out of bins and place them into matching containers.

Google Apollo 2 humanoid robot

This division of labor has something in common with UBTech's exploration: task understanding, workflow tracking and continuous action execution are completed through coordination among different models. For a robot, understanding "put the part into the container" still needs to be further translated into specific operations such as approaching the object, adjusting the grasping pose, moving and placing.

The role of general-purpose models is also extending into the robot development process.

According to Google, Gemini 4 Argon participated in the development of Gemini Robotics 2.5, and can assist in building digital twins based on real camera footage, generate simulation exercises, and analyze the causes of task failures in combination with videos, helping researchers test and improve solutions. These simulation data, combined with real operation demonstrations, are used to improve the robots' ability to adapt to new scenarios and new objects.

This offers another angle for observing the relationship between the two types of models: general-purpose models can not only participate in robots' task execution, but also help collect and generate training materials and analyze the causes of failures, accelerating the improvement of specialized models.

Therefore, between general-purpose models and embodied models, there are both changes in the scope of capability coverage and collaboration in R&D and application. After general capabilities improve, which tasks can be completed directly and which still require training through robot data is being tested through specific tests.

After models become smarter, speed and cost still matter

Whether general-purpose models can take over more robot work still requires answering deployment questions.

In NVIDIA's Astranomy research, the direct control approach took an average of about 18.56 seconds per inference; for the agent approach that achieved the highest success rate, the API cost per successful task was estimated at about $36.40. The research clearly points out that latency and cost are important constraints in practical applications.

Improvements in success rates are not always accompanied by improvements in speed and cost. Agent approaches with more context and the ability to plan and adjust continuously may complete more complex tasks, but also require more inference and model calls. UBTech's collaborative tests show that in specific tasks, delegating part of the action execution to specialized models can reduce calls to general-purpose models.

This also means that comparisons of robot models cannot rely on a single success-rate number alone. Grasping in simulation, bowl stacking on real machines, and parts kitting on automotive production lines involve different task conditions and operational requirements, and the conclusions they can support each have their own scope.

In real work environments, objects and people do not pause while the model thinks. Between a robot eventually completing a task and continuously working at the pace of a production line, there still lie response speed, operating cost and the ability to recover after errors.

From simulation tests to factories, model scores still need to be translated into specific operational metrics: whether parts can be grasped accurately after a change, whether operational deviations can be corrected automatically, how many manual interventions are needed per shift, and how much each completed operation costs. These questions will also determine how much work general-purpose models can directly undertake and which steps still require specialized training.

(Source: The Paper) Mid-article embedded advertisement

Source

East Money