OpenAI Reassigns Robotics Staff as General AI Models Show Direct Hardware Control
OpenAI is integrating its robotics teams into main research divisions, betting that general-purpose models like GPT-6 Astra can direct physical robots without specialized software.
OpenAI is once again restructuring its internal robotics operations, integrating specialized hardware and software personnel into its broader AI divisions. Rather than stepping away from physical automation, the company aims to accelerate progress by merging robotics researchers into the core research group led by Chief Research Officer Mark Chen. Concurrently, applied robotics and hardware engineers are joining the Applied organization under Greg Brockman. The decision was executed on such short notice that a booth reserved by OpenAI at the Conference on Robot Learning (CoRL) in Austin remained unoccupied, according to a report by Fortune.
General models demonstrate spatial control capabilities
This organizational shift stems from OpenAI's belief that general-purpose foundation architectures can manage physical machines directly, potentially reducing the need for dedicated robotics software. Benchmark evaluations suggest large multimodal models can outperform domain-specific designs. In a study conducted by Nvidia, an agent driven by OpenAI's GPT-6 Astra achieved a 74.8 percent success rate on simulated robotics tasks using a virtual DROID platform, compared to 51 percent attained by the top specialized models.
Further evaluations highlight this performance gap. On the StationeryBench benchmark—which evaluates dual-arm YAM robotic arms executing five desk-object manipulation tasks across 200 trials—GPT-6 Astra fully completed 7 out of 100 assigned tasks, whereas Ai2's MolmoAct2 completed none. Astra recorded an average progress score of 46, compared to 12 for MolmoAct2. In a separate test, Astra running on a humanoid robot successfully navigated and cleaned up a kitchen environment.
Academic researchers have noted these underlying spatial advances. Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, characterized Astra as a major leap in spatial understanding. Artzi observed that on the unpublished REMAP benchmark, GPT-6 Astra achieves near-human accuracy, speculating that OpenAI may have incorporated extensive 3D training data such as Blender scenes.
Strategic realignment amidst financial pressures
The move marks another chapter in OpenAI's changing approach to hardware. The organization previously shut down its robotics division in 2020, with co-founder Wojciech Zaremba citing a lack of necessary training data and noting that large language models offered a faster path to AGI. After rebuilding the team in 2024, CEO Sam Altman reportedly targeted 2027 for the company's first public robot demonstration, with long-term goals to manufacture everyday robots.
However, OpenAI currently faces financial pressure to focus on its core AI business, where it has lost ground to competitor Anthropic. At the same time, rival foundation model developers including Anthropic and Google have also logged recent progress on robotics tasks using their general-purpose models.
What it means for developers
For software engineers and hardware developers, OpenAI's reorganization highlights a shift toward using general-purpose multimodal models for high-level spatial reasoning and physical planning. If foundation models continue to generalise effectively across physical tasks, engineering teams may rely more on broad vision-language systems rather than building task-specific control models from scratch.
However, specialized robotics models maintain clear operational advantages in specific deployment scenarios. Domain-specific architectures often retain edges in execution speed, energy consumption, lower inference cost, and overall system reliability compared to large frontier models.
As general intelligence models expand into spatial reasoning and robotics control, developers can try top AI models cheaply through one API at https://apixoai.online to evaluate frontier performance across different platforms. Deciding between fine-tuned specialized control models and general multimodal models will remain a central architectural choice as robotic automation advances.
Source: OpenAI reassigns robotics researchers to other teams as general-purpose models might make specialized robotics AI unnecessary — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
