No. 120
2026-11-16
the AI pulse — brewed, not raw
 AI CRITIQUE

OpenAI dissolves its independent robotics team and redirects researchers to its general-purpose models

Read the article    English (automatic translation)

Summary

OpenAI has disbanded its independent robotics team, integrating its researchers into the general-purpose model research division led by CRO Mark Chen and the applied hardware unit under Greg Brockman. The company believes that frontier models like GPT-6 Astra can serve as the 'brain' for robots, eliminating the need for specialized robotics AI.

Why it matters

This strategic pivot directly challenges a core assumption of the multi-billion-dollar robotics startup ecosystem: that robots require dedicated, specialized AI. If general-purpose models can successfully handle manipulation and physical reasoning, this will reshape the competitive landscape for Figure, Tesla Optimus, and dozens of rivals. The timing also highlights OpenAI's financial strain: quarterly operating losses have reached $10.9 billion, with a potential IPO in 2027 on the horizon.

Highlights

  • According to an Nvidia study cited in the article, an Astra-based agent achieved a 74.8% success rate in simulated manipulation benchmarks, while the best specialized robotics model managed only 51.0%.
  • OpenAI originally disbanded its robotics team in 2020, only to re-establish it in 2024 through a partnership with Figure. This collaboration ended in early 2025.
  • The restructuring coincides with Altman's efforts to shed 'side quests' amid quarterly losses of $10.9 billion and a possible 2027 IPO.

Critical take

The article relies on a single Nvidia benchmark in a simulated environment to justify the strategic shift. Experts like Tedrake and Florence rightly point out that real-world factory data, latency constraints, and on-device inference remain unsolved problems that a cloud-dependent frontier model does not address. The empty CoRL booth is an illuminating detail; however, the article provides no independent verification of the 74.8% figure or any evidence of performance in physical, non-simulated environments.

This critique was written by the local AI model Qwen3.8-27B. I chose the story; I do not review each critique before it is published.
Back to the issue