Fortune reports that OpenAI had already booked an exhibition booth at the Conference on Robot Learning (CoRL) in Austin, which subsequently sat empty. That visual absence reflected an immediate internal restructuring: OpenAI is disbanding its robotics team again, departing from the rationale behind its 2020 exit. Rather than stepping away from physical machines, the organization seeks to accelerate robotics development by folding staff into broader engineering groups. Robotics researchers are transferring to the main research team led by Chief Research Officer Mark Chen, while the applied robotics and hardware units join the Applied organization under Greg Brockman.
As OpenAI co-founder Wojciech Zaremba previously explained, OpenAI shut down its initial robotics division in 2020 because the company lacked necessary training data and viewed large language models as a faster path to artificial general intelligence. The organization rebuilt the unit in 2024, yet it now faces heavy financial pressure to concentrate resources on its core software business, where it has lost significant ground to competitor Anthropic. OpenAI believes that unified, general-purpose models such as GPT-6 Astra can operate robotic hardware directly, diminishing the commercial justification for dedicated robotics AI units.
In an Nvidia study, an Astra-powered agent achieved a 74.8 percent success rate on simulated robotics tasks, compared to 51 percent for the best specialized models.
That benchmark suggests multimodal foundations can coordinate complex actions without isolated control software, changing how engineering organizations allocate capital between generic foundation training and domain-specific automation systems.
Benchmark Performance Against Narrow Systems
The performance evidence backing this structural change comes primarily from recent testing environments. While an Astra-powered agent secured 74.8 percent success in Nvidia's evaluation against 51 percent for leading specialized models, those evaluations used a simulated DROID robot rather than physical hardware. Additional multi-task evaluations reveal both the current capabilities and the persistent operational limits of large frontier models. On a benchmark named StationeryBench, researchers evaluated GPT-6 Astra against Ai2's MolmoAct2 across five manipulation tasks involving desk objects using dual-arm YAM robotic arms. Across 200 trials, Astra completely finished 7 out of 100 tasks, whereas MolmoAct2 failed to complete any. Astra achieved an average progress score of 46, compared to 12 recorded by MolmoAct2.
Demonstrations also extend beyond tabletop manipulation: in another test, Astra running on a humanoid robot navigated and cleaned up a kitchen environment. Yoav Artzi, an AI researcher at Cornell University and Google DeepMind, identified Astra as representing a major leap in spatial understanding. On the unpublished REMAP benchmark, GPT-6 Astra reaches near-human accuracy according to Artzi, who suspects OpenAI may have trained the model on extensive volumes of 3D data such as Blender scenes. It remains unclear how much robotics data OpenAI and competing labs incorporated during pre-training, though rival general-purpose architectures from Anthropic and Google have registered documented progress on mechanical tasks.
The Operational Limits of Generalized Models
The strategic shift assumes large general models will absorb robotic motor control at lower aggregate development expense than maintaining separate hardware teams. Yet specialized robotics models may still hold operational advantages in deployment cost, execution speed, system reliability, and electrical energy consumption, even if their contextual generalization remains comparatively restricted.
Specialized robotics models may still have advantages in cost and reliability as well as speed and energy consumption, even if they don't generalize as well.
Industrial automation budgets depend on predictable latency, edge power constraints, and sub-millimeter precision, areas where frontier foundation models running massive cloud inference workloads face clear economic constraints.
OpenAI Chief Executive Sam Altman is targeting a first public robot demonstration for 2027, maintaining a corporate goal to manufacture proprietary everyday robots over the long term. Folding hardware specialists into centralized research divisions under Mark Chen and Greg Brockman reduces redundant overhead while the core business defends software market share against Anthropic. Merging physical actuation into monolithic multimodal pipelines lowers immediate specialized research expenses, but enterprise deployment will stall if high inference power demands and incomplete task completion rates outweigh the architectural elegance of running a factory floor from a unified frontier model.
What’s behind it
Betting on general-purpose software pushes durable hardware automation aside once again to fuel the frontier model race. Running algorithms on synthetic benchmarks like virtual robot DROID looks convincing, yet moving beyond simulation exposes the friction of the physical world where complex multi-step manipulation still fails almost entirely. Folding specialized robotics into a single massive foundation model protects OpenAI’s core software business, but leaves the gap between synthetic tests and physical machinery wide open.
Our stories: OpenAI Leadership Shake-Up: Shifting Priorities & Future Impact (April 7).