OpenAI has integrated its robotics team into its core research division, citing the ability of its GPT-6 Astra model to handle robot control tasks without specialized models. The move, reported by Fortune on November 13, follows internal disputes over its scope and the practical limits of general-purpose AI in robotics.
TL;DR
- OpenAI reportedly folded its robotics team into core research, citing GPT-6 Astra's ability to control robots without specialized models.
- Staff disputed the report, with engineers calling robotics work 'alive and well' and pointing to 15 open roles.
- NVIDIA's October 14 tests showed Astra achieved 66.3% success in simulated robot tasks, but with pauses and delays that raise questions about real-world reliability.
What happened
OpenAI moved its robotics team into its main research and product groups, according to a Fortune report published November 13. The company stated that models like GPT-6 Astra can run robots without requiring a separate, dedicated robotics model.
The reported shift faced pushback from employees. Engineer David Schnurr called the robotics work 'alive and well' and highlighted 15 open roles, while another employee labeled the report 'completely wrong.' The scope and impact of the reorganization remain unclear.
NVIDIA researchers tested GPT-6 Astra on simulated robot tasks in an October 14 report. In the RoboLab-120 benchmark, Astra achieved 66.3% success when generating actions directly, though the setup included robot control software and paused the simulation while the model worked.
Google DeepMind's November 10 announcement of Gemini Robotics 2.5 reported 96% success in a trial picking and packing automotive parts using Apptronik's Apollo 2 on Mercedes-Benz production lines. The system combines a workflow model with Gemini Robotics 2.5 for actions, leaving questions about productivity and safety in real-world settings.
Why it matters
The reported integration suggests OpenAI is prioritizing general-purpose models like GPT-6 Astra for robotics, potentially reducing the need for specialized robotics models. This shift could reorient research toward improving shared intelligence that benefits multiple domains, including robotics.
For robotics researchers, the move implies that advances in general models could directly enhance robot behavior, offering a unified approach to intelligence across tasks. However, NVIDIA's experiment highlights challenges: pauses in simulation and average action generation times of 18.56 seconds raise concerns about real-time responsiveness in dynamic environments.
For teams deploying robots, the location where models run—cloud-based versus local—can significantly impact performance. Google's June 2025 announcement of a local robotics model emphasized fast, offline responses, suggesting a role for specialized models even as general reasoning improves.
Hardware engineers may also benefit from better general models, which could assist in design, simulations, and reviews. OpenAI noted on November 3 that its hardware teams, including robotics, use ChatGPT and Codex for tasks like manufacturing drawings and design reviews. Yet, OpenAI's own example shows that models can miss critical manufacturing constraints, requiring human oversight.
Key facts
- OpenAI moved its robotics team into core research, per Fortune's November 13 report.
- The company cited GPT-6 Astra's ability to control robots without specialized models.
- NVIDIA's October 14 report tested Astra on RoboLab-120, achieving 66.3% success in simulated tasks.
- NVIDIA's simulation paused while Astra generated actions, with an average time of 18.56 seconds per action.
- Google DeepMind's November 10 announcement reported 96% success for Gemini Robotics 2.5 in a Mercedes-Benz parts-picking trial.
- OpenAI engineer David Schnurr stated robotics work is 'alive and well' and pointed to 15 open roles.
- OpenAI said on November 3 that hardware teams use ChatGPT and Codex for design and manufacturing tasks.
- Google announced a local robotics model in June 2025, emphasizing offline, fast responses.
Context
OpenAI's shift reflects a broader trend of leveraging general-purpose AI models for specialized tasks, reducing reliance on domain-specific systems. This approach aligns with the company's focus on scaling foundational models that can adapt across applications, including robotics.
Robotics research has historically relied on a mix of specialized models and hardware-specific solutions. Google DeepMind's pilot with separate reasoning and action models shows an alternative path, where workflow management and action execution are decoupled.
The practical limits of general models in robotics remain a key debate. While NVIDIA's simulation results suggest potential, real-world deployment requires addressing real-time constraints, reliability, and safety—areas where specialized models may still play a critical role.
