OpenAI Disbands Robotics Team, Pivots to General Models for Embodied AI

• OpenAI disbanded its independent robotics team, moving researchers to the main research team led by Mark Chen and hardware staff to the Applied organization under Greg Brockman.
• The strategic shift relies on general-purpose frontier models, specifically GPT-6 Astra, to control robots directly, eliminating the need for dedicated robot-specific AI models.
• Nvidia research indicates a GPT-6 Astra-based agent achieved a 74.8% success rate in simulation manipulation benchmarks, outperforming the top dedicated robot model at 51.0%.
• The move aligns with cost-cutting pressures and a focus on core products, with OpenAI targeting a first humanoid robot demo in 2027 and preparing for an IPO in early 2027.

Strategic Context and Limitations

• Sam Altman directed the abandonment of "side quests" to reallocate computing resources to core, profitable businesses, contributing to the team dissolution.
• Experts note that while general models reduce the need for specialized training, LLMs are slow and require cloud hosting, necessitating constant network connectivity for robots.
• MIT professor Russ Tedrake argues that robot startups retain an advantage through real-world deployment data, which becomes more valuable as foundation models improve.