AI Research

OpenAI Disbands Standalone Robotics Team — Researchers Folded Into Main Org, Betting on GPT-6 Astra for Embodied AI

OpenAI robotics team reorganization with GPT-6 Astra model controlling a humanoid robot

OpenAI Disbands Standalone Robotics Team — Researchers Folded Into Main Org, Betting on GPT-6 Astra for Embodied AI

OpenAI Disbands Robotics team on November 13, 2026, moving researchers into its core research organization (under Chief Research Officer Mark Chen) and applied/hardware teams into the Applied organization (under President Greg Brockman). The company confirmed the reorganization to Fortune — not a retreat from robotics, but a strategic bet that general-purpose models (GPT-6 Astra) can now control robots directly, eliminating the need for specialized robotics models.

This is OpenAI's second robotics wind-down — first in 2020 (post-Rubik's Cube), then rebuilt in 2024 with Figure AI partnership (ended Feb 2025 when Figure developed its own Helix model). Now, GPT-6 Astra's demonstrated robotics capability triggers the fold-in.

The Reorganization: Where Researchers Went

Former Team New Home Leadership Focus
Robotics Research Main Research Org Mark Chen (CRO) Core robotics science, simulation, model integration
Applied Robotics Applied Org Greg Brockman (President) Real-world deployment, hardware integration
Hardware Applied Org Greg Brockman Robot hardware, sensors, actuators

No layoffs — researchers reassigned. OpenAI booth at CoRL (Conference on Robot Learning, Austin, Nov 9–12) sat empty — registered but no display/demo. Former team members present but no official presence.

The Strategic Thesis: "General Models Run Robots"

Core bet: GPT-6 Astra and successors can directly control robots — no specialized robotics model needed.

Evidence Detail
Sora's physics understanding Video generation required implicit world model; transferred to robotics
GPT-6 Astra robotics benchmarks Directly applied to robotic tasks; "surprising performance" (Fortune)
NVIDIA simulation study GPT-6 Astra beat every specialist robot policy on accuracy (TechTimes)
Latency caveat Each Astra call: 11–19 seconds — needs fast controller underneath
Industry shift CMU's Guanya Shi: OpenAI treats robotics as "downstream application of general intelligence"

OpenAI's statement: "At the start of the year, researchers assumed best path was building on Sora (physics understanding). Since then, general frontier models including GPT-6 Astra have shown they can be applied directly to robotics. This discovery motivated the reorganization."

The Figure AI Partnership: Ended, Lesson Learned

Phase Timeline Outcome
2020 First robotics wind-down Post-Rubik's Cube; limited data/hardware
2024 Rebuild + Figure AI partnership Co-develop models for Figure humanoids
Feb 2025 Figure ends partnership Figure builds own Helix model (System 1/2 design)
2026 GPT-6 Astra proves capable Standalone team folded into core

Figure AI's path: Built Helix (System 1 fast control + System 2 planning); BMW Spartanburg deployment — 90K+ parts moved, 1,250+ hours on Figure 02. Figure CEO Brett Adcock: "Hard problem is high-rate robot control" — needs fast controller layer.

Competitive Contrast: Rivals Expanding Dedicated Robotics

Company Robotics Strategy
OpenAI Fold into general models (GPT-6 Astra → robots)
Google DeepMind Expanding dedicated robotics (Gemini Robotics, AlphaGenome 2)
NVIDIA Simulation-first (Isaac Sim, Project GR00T, fast controllers)
Tesla End-to-end (FSD stack → Optimus; real-world data flywheel)
Figure AI In-house Helix (System 1/2; BMW deployment)
Generalist (startup) General-purpose models for robots (Pete Florence: "unclear which model dominates")

OpenAI's contrarian bet: Software-first, general-model approach vs. rivals' hardware-data co-development.

The 2027 Humanoid Demo Target — Unchanged

Target Detail
Public humanoid demo 2027 (Sam Altman target, unchanged)
Approach GPT-6 Astra (or successor) as brain; fast controller underneath
Hardware In-house development (post-Figure)
Cost pressure $10.9B Q3 operating loss; "abandon side quests" (Altman)

Cost context: Q3 operating loss $10.9B (AIPress Post 94); Altman told company to "abandon side quests" — robotics team as "side quest" despite long-term vision.

The Latency Problem: 11–19 Seconds per Astra Call

Metric Value Implication
Astra inference per call 11–19 seconds Too slow for real-time control
Needed Fast controller (System 1) High-rate motor commands
Architecture System 2 (Astra) → System 1 (controller) Figure's Helix mirrors this
NVIDIA study Astra beats specialists on accuracy But speed gap remains

Implication: General model = planner (System 2); needs dedicated fast controller (System 1) for real-time. OpenAI may build or buy this layer.

Industry Reaction: "Robotics No Longer a Special Problem"

Voice Take
CMU Guanya Shi OpenAI sees robotics as "downstream application of general intelligence" — like coding/legal tools
Pete Florence (Generalist) "Unclear any existing model dominates robotics long-run"
Fortune framing Company believes general models can start running robots without dedicated team
The Decoder update OpenAI clarified: "reorganized to work more closely with research/product; robotics work continues"

What Comes Next

  1. Integration velocity — How fast does robotics research merge into main model training? (GPT-6.5/7 robotics benchmarks)
  2. Fast controller development — Build or buy System 1 layer; acquire robotics control startup?
  3. Hardware path — In-house humanoid (2027 demo target); no Figure partnership
  4. Simulation → real transfer — NVIDIA Isaac Sim / Isaac Lab as bridge
  5. Talent retention — Robotics researchers in general research: do they stay or leave for DeepMind/Figure/Tesla?
  6. 2027 demo — What does the humanoid actually do? Teleop? Autonomous? Partial?
  7. Cost discipline — $10.9B loss drives focus; robotics must show path to revenue

The Bigger Picture

OpenAI Disbands Robotics team move crystallizes the industry's central debate: General intelligence vs. specialized embodiment.

  • OpenAI's thesis: AGI solves robotics as a downstream task. Invest in the general model; robotics follows.
  • Rivals' thesis: Embodiment requires specialized data, hardware, control loops. General models lack high-rate control, real-world physics data, hardware co-design.

If OpenAI is right: Robotics becomes a software feature of frontier models — massive leverage, lower cost. If rivals are right: OpenAI cedes the physical AI race to DeepMind, Tesla, Figure, NVIDIA — who own the data flywheel (real-world deployment → data → better models).

The 2027 humanoid demo is the proof point. Until then, the empty CoRL booth symbolizes a bold, risky bet — that the path to embodied AI runs through language models, not robotics labs.


Related AIPress Coverage

Sources

  • Fortune (Nov 13, 2026): Exclusive "OpenAI disbands stand-alone robotics team as general AI models start running robots"
  • The Decoder (Nov 13/14, 2026): Reorganization details, OpenAI clarification
  • Science Monitor (Nov 13, 2026): Physical AI race shift analysis
  • Seeking Alpha (Nov 13/14, 2026): Reorganization report
  • Startup Fortune (Nov 13, 2026): Team restructure details
  • The Robotics Media (Nov 13, 2026): GPT-6 Astra focus
  • TechTimes (Nov 16, 2026): NVIDIA simulation study, 11-19s latency
  • Humanoid Guide (Nov 2026): Humanoid work continues
  • CMU Robotics Institute / Guanya Shi comments (Fortune)
  • Generalist / Pete Florence comments (Fortune)
  • BMW / Figure AI deployment data (2025–2026)
  • NVIDIA simulation study (arXiv 2609.38537 referenced)

Building something with AI?

DevsIsle designs and ships AI systems, agents and integrations for teams that need it done properly.

Talk to our team →