OpenAI has disbanded its separate robotics division, citing a change in concept. According to an OpenAI representative, the company does not abandon the goal of creating general-purpose robots, but does not see the need to develop highly specialized AI models intended only for robots, since universal models can easily perform these functions.
Image Source: Steve A Johnson/unsplash.com
According to an OpenAI representative, the versatile GPT-6 Astra model has shown unexpected success in performing selected robotic tasks. Meanwhile, at the beginning of the year, the company’s researchers assumed that the development of this area would be based on the developments of the Sora video generator, capable of simulating the physical world.
But then it turned out that OpenAI’s versatile, advanced models could be directly applied to robotics. That is why the separate robotics team was disbanded, and the specialists working in it were transferred to groups working on more universal AI models.
The robotics research group became part of the core research division under the leadership of research director Mark Chen, and the applied robotics and hardware specialists were transferred to the applied development division, which reports to OpenAI President Greg Brockman.
This approach could challenge the reliance of many robotics startups on specialized models with “spatial intelligence.” However, practical experience in operating robots and data collected in real conditions retain their own value.
Russ Tedrake, co-founder and CEO of Walden Robotics, noted that robotics companies have an advantage that AI labs don’t: Their robots already get hands-on experience in real-world environments, such as on the factory floor.
Better base models increase the value of object-specific data because they allow robots to learn from it faster, he said.
Universal models are already being used to control robots without additional training using robotic data. In such experiments, the model receives images from cameras, information about the robot, and access to an interface through which it can issue commands to it.
In the Nvidia study, the GPT-6 Astra-based agent successfully completed 74.8% of object manipulation tasks. In comparison, the best published result for specialized robotics models in the same ranking was 51%. However, the tests took place in a simulation: the model controlled a virtual robot, so these results cannot yet be transferred to real devices.
Large universal models also have limitations: processing requests takes time, and using a cloud service requires a constant connection to the network. Therefore, Guanya Shi, an assistant professor at the Robotics Institute at Carnegie Mellon University, believes that compact, specialized models are still needed – they can work directly on the robot’s on-board computer.
Which approach will be more successful in the long term is not yet clear, said Pete Florence, co-founder of Generalist, a startup that develops universal AI models for robots.
If you notice an error, select it with the mouse and press CTRL+ENTER.
