OpenAI integrates robotics researchers into core teams
OpenAI is shifting its robotics strategy to use general-purpose models like GPT-6 Astra instead of building specialized robot AI.
The short version
- OpenAI is moving robotics researchers into the main research team and the applied robotics/hardware organization.
- GPT-6 Astra achieved a 74.8 percent success rate in simulated tasks, beating the 51 percent rate of specialized models.
- In a StationeryBench test, Astra completed 7 out of 100 tasks involving two-armed robot arms, while a competitor completed zero.
- The shift suggests that general models may provide better spatial understanding than purpose-built robotics AI.
- OpenAI aims for a public robot demo by 2027 and plans to build its own everyday robots.
Why it matters
This shift indicates that the future of robotics may rely on 'general intelligence' rather than niche software. For users, this could mean more versatile robots that can handle diverse household tasks without needing specific programming for every action.
The path toward everyday robots
OpenAI's long-term goal involves the creation of its own everyday robots. CEO Sam Altman has reportedly targeted 2027 as the year for the first public demonstration of such a robot. This aligns with the company's broader ambition to move beyond software and into physical hardware that can interact with human environments.
This direction is also influenced by the company's current financial situation. OpenAI faces significant pressure to focus on its core AI business, where it has recently lost ground to competitors like Anthropic. By integrating robotics into its main research pipeline, OpenAI can leverage its existing infrastructure to develop physical products more efficiently.
The company's history with robotics is complex, as it had previously closed its robotics department in 2020. At that time, co-founder Wojciech Zaremba stated that the lack of training data made large language models a faster path to AGI. By rebuilding the team in 2024 and now integrating it into the core research, OpenAI is betting that general models have finally bridged that data gap.
What to watch next
The next major milestone for OpenAI will be the public debut of its robot technology. Investors and the public will be looking for how well the GPT-6 Astra model translates from simulated environments to real-world physical hardware. The success of this transition will determine if general models can truly handle the unpredictability of human homes.
Additionally, the competition from other tech giants will be a key factor to monitor. Both Google Deepmind and Anthropic have recently reported progress in robotics tasks using their own large models. The race is now to see which company can produce a reliable, cost-effective robot that can perform complex household chores autonomously.
Finally, the specific types of data used to train these models remain a point of interest. While it is suspected that 3D training data is a major component, the exact scale and variety of these datasets will be crucial for achieving human-level accuracy in spatial navigation. Researchers will be watching for upcoming benchmarks like REMAP to see how these models compare to human performance.
Try it
You can look for upcoming news regarding the 2027 robot demo from OpenAI to see how their general-purpose models perform in real-world physical environments.
Watch
Will specialized models eventually regain an edge in cost-efficiency for industrial use?
How will OpenAI balance its focus on core AI business with the high costs of robotics hardware?
What specific 3D datasets will be used to reach human-level spatial accuracy?
OpenAI integrates robotics researchers into core teams
General models outperform specialized AI
The primary reason for this internal restructuring is that OpenAI's general-purpose models, such as GPT-6 Astra, appear capable of controlling robots directly. This capability suggests that the company may no longer need to develop separate, specialized robot AI models to achieve complex movements or task completion.
Evidence for this shift comes from a study conducted by Nvidia. In this research, an agent powered by Astra achieved a 74.8 percent success rate in simulated robot tasks. This performance significantly outperformed the best specialized models, which reached only a 51 percent success rate during the same tests.
Benchmarks and training techniques
Further testing on the StationeryBench highlighted the gap between general and specialized systems. When tasked with handling desk items using two-armed robot arms, Astra completed 7 out of 100 attempts fully. In contrast, the MolmoAct2 model failed to complete a single task out of 200 attempts, while Astra maintained a much higher average progress score.
Experts believe these results stem from how the models are trained. AI researcher Yoav Artzi suggests that OpenAI might be using massive amounts of 3D training data, such as Blender scenes, to improve the models. This approach helps the AI develop a spatial understanding that allows it to navigate and interact with physical objects more effectively.
Sources and links
- Officialeiner Nvidia-Studie research.nvidia.com
- Officialanthropic.com/research/claude-plays-robotics anthropic.com
- Postsieht in Astra ebenfalls einen klaren Sprung beim räumlichen Verständnis x.com
- Postx.com/yoavartzi/status/2098511167090000086 x.com
- PaperGeschwindigkeit und Energieverbrauch arxiv.org
- PaperKosten und Zuverlässigkeit doi.org
- ReportingThe Decoder: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by The Decoder · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
