Study says GPT-6 Astra's robot skills may come from training on one robot arm

The demos that set off the debate

On September 3, 2026, OpenAI released GPT-6 Astra, a general-purpose model that handles text and images. Its robotics performance quickly drew attention. NVIDIA brought it into its robot simulation and design tools.

In one demo, developers paired Astra with PTC Onshape and NVIDIA Isaac Sim. They built a 3D model of a car suspension, adjusted a wrench so it could reach the bolts, and ran a simulated job in which a robot removed parts with it.

In another, a Unitree G1 humanoid robot walked through a room it had never seen. It remembered where objects were and tidied up after only vague spoken instructions. Some people in robotics went as far as declaring that a whole earlier approach to robot control was finished.

That raised a question. Do these skills appear on their own once a model gets general enough, or did the model absorb a lot of specialized robotics knowledge in training? As BigGo Finance reported, OpenAI's blog does not say what went into the training data, or whether robotics data was part of it.

How the researchers probed the models

A new study tried to get at the answer indirectly. The researchers gave 50 open and closed models two kinds of robot-arm math problems. One used the specifications of the Franka Panda arm. The other used synthetic robot specifications made up for the test.

The problems were forward kinematics and inverse kinematics. In plain terms, forward kinematics means: given the angles of each joint, where does the robot's hand end up? Inverse kinematics runs the other way: given where you want the hand, what joint angles get it there? The models had to answer directly, with no outside tools, no running code and no step-by-step working.

Most of the 50 models behaved alike. Three did not: GPT-6 Astra, GPT-6.1 Sol and Gemini Argon. The researchers say these three differed markedly from the rest, and they argue that GPT-6.1 Sol and Gemini Argon received similar training to Astra.

The swap test that raised suspicion

The key step was a control experiment. The researchers changed the robot's DH parameters. These are a standard set of numbers describing how a robot arm's joints and links relate geometrically, including link length, twist, offset and joint angle. Change the numbers and you have described a different machine.

A model that understands the math should simply recalculate. The researchers found that even when the three models were handed the DH parameters of a different seven-degree-of-freedom robot, their answers still looked like the Franka Panda's.

The authors read this as overfitting to one robot body. Put simply, the models may have memorized how that particular arm moves instead of learning the general rules that work on any arm. The distinction matters because the two can look identical on a familiar robot and only come apart on an unfamiliar one.

What the study does not show

The source is careful here, and so should readers be. The study is described as preliminary. It cannot show that Astra's robot abilities come entirely from specialized training, and it cannot identify which robotics data OpenAI or Google used. The source also includes no response from OpenAI.

The researchers released their full results, code and model inference records so others can reproduce them. Some researchers said they saw a similar familiarity bias toward the Franka Panda arm in their own tests. Others suggested running the same tests on the Unitree G1 humanoid. The authors say they plan to add more robot types and study how specific training datasets affect results, but that this needs more computing budget, measured in tokens.

Note also that the test measures one narrow skill: arm math answered without tools. It does not directly measure the navigation and tidying shown in the humanoid demo.

Why markets and builders are watching

The debate lands while the robotics sector is rallying despite a shaky broader market. The source says market commentators credit progress in connecting large models to physical action, not just language. Generalization, meaning working when the scene changes, is the main limit on robot usefulness, and weak generalization is the biggest challenge for humanoid robots today.

The same commentators concede that robot "brains" still think slowly, use many tokens and need better accuracy. They also note that the humanoid robot index's price-to-earnings ratio sits in the lower-middle part of its historical range. Their advice is "positive direction, patience first."

One practical detail: the source says single purchases of real-world robot data are typically around 100,000 hours, with next year expected to reach a million hours. If general models still lean on specialized data, that demand makes sense. Whether robot skills emerge or are taught, the training data question now sits at the center of the debate.