AI News

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

Sina TechUpdated Nov 14 17:47

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. The report also summarizes several tests: NVIDIA Astranomy research shows that an agent solution based on GPT-6 Astra achieved a 74.8% success rate on the RoboLab-120 simulation benchmark, higher than the 51.0% of the specialized model Lakeside; in six sets of UBTECH tests, Astra failed to directly grasp a Rubik's cube, simulate carrying a basket, and directly stack bowls, but after coordinating with a self-developed embodied model it completed lucky-bag grasping, obstacle avoidance, and bowl stacking, with token consumption and estimated API costs in the bowl-stacking task reduced by 83.6% and 74.2%, respectively. The NVIDIA research also noted that the direct-control approach takes about 18.56 seconds per inference on average, and the highest-success-rate approach costs an estimated about $36.40 in API fees per successful task; latency and cost are important limits on practical application.

Why read this

AI score 71

Simulation and real-robot tests from multiple institutions are presented side by side, allowing readers to clearly see the capability boundaries and cost constraints of general-purpose models on robotic tasks.

Topics

Tags

Body · AI translation

Related topics top start Related topics top end Futures promotion begin Futures promotion end Miaopai begin Miaopai end Video player start Video player end Market chart begin Market chart end

The Paper reporter Qin Sheng

Progress in the capabilities of general-purpose large models on robots 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 became an important milestone: the capabilities it demonstrated have led the industry to pay closer attention to 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 had disbanded its standalone robotics team, with robotics researchers folded into the core research team and the applied robotics and hardware teams moved into the applications division. 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 to real robots, 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 approach based 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, 74.8% corresponds to the agent approach with persistent interaction context; the approach that directly generates actions achieved a 66.3% success rate. The tests mainly covered manipulation 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 it cover complex contact tasks such as insertion and assembly. NVIDIA test results

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

Anthropic's Claude Fable 5.2 and Mythos 5.2, in the RoboDojo Sim dual-arm robot simulation test run internally by the company, Fable 5.2 scored 64.9%, while under the same test framework GPT-6 Astra scored 29.8%. 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 NVIDIA's research, so the scores cannot be directly compared across tests.

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 a breakthrough may occur in the next one to three years. He demonstrated on site Unitree's robot tests 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-robot and simulation tests around GPT-6 Astra, covering real-robot grasping of a Rubik's cube, grasping a lucky bag, and autonomous obstacle avoidance, as well as basket carrying in a simulated environment, direct control for stacking bowls, and bowl stacking combined with a world model.

According to results disclosed by the company, Astra's direct grasping of the Rubik's cube, simulated basket carrying, and direct bowl stacking were all incomplete; 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's Astra Token consumption and estimated API cost were reduced by 83.6% and 74.2%, respectively. UBTech tests Astra direct control for grasping a Rubik's cube

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

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 robot system is still improving practical operating capabilities through the combination 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 manipulation of flexible objects, including using tongs to pick up and place food, tying a bow on a gift box, and transferring 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 both 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, process tracking, and continuous action execution are completed collaboratively by different models. For a robot, understanding "put the part into the container" also needs to be further translated into specific operations such as approaching the object, adjusting the grasping posture, moving, and placing.

The role of general-purpose models is also extending into the robot R&D 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, generating simulation exercises, and analyzing the causes of failures in combination with videos of failed tasks, helping researchers test and improve solutions. These simulation data, combined with real operation demonstrations, are used to improve the robot's ability to adapt to new scenarios and new objects.

This provides another angle for observing the relationship between the two types of models: general-purpose models can not only participate in robots executing tasks, but also help collect and generate training materials, analyze the causes of failures, and accelerate 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 need to be trained through robot data are 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 estimated API cost per successful task was about $36.40. The research clearly points out that latency and cost are important limits for practical application.

Improvements in success rates are not always accompanied by improvements in speed and cost. Agent approaches with more context and the ability to continuously plan and adjust may complete more complex tasks, but they 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 just one success-rate number. Grasping in simulation, bowl stacking on a real machine, and parts kitting on an automotive production line 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 will not pause while the model is thinking. Between a robot ultimately completing one task and continuously working at the pace of a production line, there are still response speed, operating cost, and the ability to recover after errors.

From simulation testing to factories, model scores still need to be translated into specific operational metrics: whether it can grasp accurately after parts are changed, whether it can correct operational deviations on its own, 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.

This issue edited by Lu Feng news_keyword_pub,stock,sz300024&sh688836 iframe below the main text iframe below the main text general statement general statement QR code at the end start QR code at the end start

QR code at the end start QR code at the end start editor name and work code

editor name and work code

Source

Sina Tech