A sudden structural overhaul inside OpenAI has dismantled the company’s standalone robotics division, marking the second time in six years that the artificial intelligence developer has walked away from maintaining an isolated unit dedicated to physical machines.
The restructuring, confirmed on November 13, reallocates OpenAI’s robotics workforce across two distinct operational pillars. The core robotics research contingent has been absorbed into the centralized research organization led by Chief Research Officer Mark Chen. Meanwhile, hardware engineering and applied robotics groups have been reassigned to the Applied organization under OpenAI President Greg Brockman.
The abruptness of the transition reverberated through the academic and industrial robotics community at the Conference on Robot Learning (CoRL) in Austin, Texas. OpenAI was registered on the official exhibitor map at booth #208. When attendees arrived at the Austin convention floor, the booth sat completely empty—unstaffed, devoid of corporate signage, and lacking physical hardware demonstrations. A handful of former team researchers attended the conference as individuals, answering questions from peers while corporate representatives confirmed the organizational shift behind closed doors.
OPENAI ROBOTICS REORGANIZATION
[ Former Standalone Robotics Team ]
│
┌──────────────┴──────────────┐
▼ ▼
[ Mark Chen's Research Org ] [ Greg Brockman's Applied Org ]
• Frontier foundation models • Physical hardware integration
• GPT-6 Astra direct policy • Applied robotics engineering
• Cross-modal transfer • 2027 Humanoid Prototype track
An OpenAI spokesperson maintained that the decision does not constitute a retreat from the physical domain, stating that robotics remains integral to the company's long-term push toward artificial general intelligence. Rather, the company argues that recent leaps in frontier multimodal models have rendered purpose-built, domain-isolated robotic software stacks redundant. OpenAI’s current leadership asserts that physical machines no longer require specialized brains. Instead, they claim, massive general-purpose foundation systems can now be embedded directly into physical hardware mechanisms.
The decision to dissolve the dedicated OpenAI robotics team reflects a strategic calculation with broad implications for hardware startups, foundational research labs, and prospective investors. By treating embodiment not as a distinct branch of artificial intelligence, but as an applied downstream endpoint of centralized models, OpenAI is betting that the path to operational humanoid robotics runs entirely through large-scale compute clusters rather than specialized mechanical labs.
Yet that conceptual pivot faces immense technical skepticism, financial pressures ahead of a projected 2027 initial public offering, and an unresolved performance gap between simulated benchmarks and physical reality.
Dismantling the Silo: Frontier Models as Hardware Drivers
The operational justification for disbanding the unit stems from an internal reassessment of model generalizability that unfolded over the past year. Entering 2026, researchers inside OpenAI operated under the working assumption that embodied intelligence would require specialized world-model architectures, primarily derived from video-generation pipelines like Sora. The premise held that systems capable of photorealistically predicting physical frames would naturally encode Newton’s laws, spatial permanence, and fluid dynamics, serving as the foundational perceptual engine for robotic manipulation.
Internal evaluations shifted focus when research teams began deploying OpenAI’s newest frontier model class, designated GPT-6 Astra, directly into motor-control frameworks. Astra demonstrated an unanticipated capacity to bypass intermediate trajectory planning layers, mapping streaming multimodal inputs—high-definition stereo video, point clouds, force-torque sensor telemetry—directly into joint velocity commands and end-effector coordinate vectors.
OpenAI leadership concluded that running an insular robotics unit had created unnecessary technical friction. By walling off roboticists in their own division, research insights were insulated from the primary pre-training pipelines powering frontier models. Folding the algorithmic staff directly into Mark Chen’s core research pool ensures that every iteration of the company’s frontier foundation weights natively incorporates physical manipulation objectives during pre-training, rather than treating robotic control as an external fine-tuning exercise.
In the Applied wing under Brockman, the mandate shifts entirely toward cross-platform integration. Rather than developing closed, proprietary robotic platforms isolated from third-party manufacturers, OpenAI’s applied engineering teams are constructing standard application programming interfaces (APIs) and unified middleware designed to pipe general model intelligence into third-party chassis, wheeled mobile manipulators, industrial arms, and bipedal humanoids.
The company is treating physical hardware much as it treated web and mobile operating systems during the early rollout of ChatGPT: as passive terminal infrastructure awaiting intelligence from a centralized network.
The Simulation Paradox: Astra’s Benchmark Dominance vs. Industrial Realities
OpenAI’s decision to dissolve its standalone robotics unit was accelerated by an internal proof point derived from an extensive simulation study conducted in partnership with Nvidia.
The benchmark tasked various artificial intelligence architectures with executing zero-shot and few-shot robotic manipulation tasks inside a high-fidelity synthetic physics simulator. The tasks required complex spatial reasoning, dynamic multi-object sorting, tool assembly, and contact-rich obstacle negotiation.
SIMULATED MANIPULATION SUCCESS RATE (Nvidia Benchmark)
GPT-6 Astra (General Model) [████████████████████████] 74.8%
Top Specialized Robotics Model [████████████████] 51.0%
Simulation Performance Delta: +23.8% in favor of Astra
On the simulated benchmark, GPT-6 Astra achieved a 74.8% task completion rate. In contrast, the most capable specialized robotics baseline models—architectures trained exclusively on physical robot kinematics, human teleoperation data, and robotic domain trajectories—maxed out at a 51.0% success rate.
The 23.8 percentage-point margin convinced OpenAI leadership that general-purpose frontier architectures possess superior visual reasoning, commonsense semantic understanding, and long-horizon recovery behaviors compared to models that see only specialized robotic datasets.
THE INFERENCE LATENCY & COST GAP (Physical Deployment)
Metric GPT-6 Astra (Direct) Industrial Requirement
─────────────────────────────────────────────────────────────────────
Cost Per Attempt $36.40 in tokens <$0.001 per operation
Model Call Latency 18.56 seconds <0.020 seconds (50 Hz)
Physical Success 7 out of 100 tasks 99.9% (Six Sigma)
However, data emerging from physical hardware deployments reveals a stark divergence from simulation.
When researchers ported Astra from the frictionless environment of synthetic physics engines to real physical robotic arms operating in unstructured laboratory environments, the system completed only 7 out of 100 manipulation tasks.
The failure points highlight the deep divide between high-level conceptual reasoning and the sub-millisecond physical feedback loops that govern real-world contact dynamics:
- Inference Latency Bottlenecks: Astra averaged 18.56 seconds per model call. In an industrial or collaborative home setting, operational control loops must cycle at minimum rates of 20 to 50 Hertz (20 to 50 milliseconds) for kinematic trajectory tracking, and between 500 Hertz and 1 Kilohertz for impedance and force control. An 18-second pause between observations leaves a physical robot blind to shifting environments, dropped objects, and dynamic human movement.
- Prohibitive Unit Economics: Deploying Astra directly as an end-to-end motor controller consumed an estimated $36.40 in compute token costs per pick-and-place attempt. While acceptable for academic simulation proofs, these economics are untenable for logistics warehouses, manufacturing floors, or consumer domestic assistance, where individual item transactions must cost fractions of a cent.
- The Sim-to-Real Perceptual Gap: Real-world tactile friction, subtle motor backlash, non-rigid object deformation, and sensor glare regularly disrupted Astra’s policy execution, leading to dropped items, stalled actuators, or unrecoverable positional drift.
By embedding the robotics researchers directly into Mark Chen’s broader model optimization groups, OpenAI hopes to address these specific latency and cost failure modes through architectural quantization, specialized speculative decoding pipelines for action tokens, and hierarchical distillation.
A Six-Year Structural Odyssey: 2020, 2024, and the Second Reorganization
This latest organizational shift represents OpenAI's third structural reboot in robotics since the laboratory's founding. The company’s volatile relationship with physical machines illustrates an ongoing internal debate over whether embodiment is an essential catalyst for artificial general intelligence or a compute-draining diversion.
OPENAI ROBOTICS TIMELINE: A RECURRING STRUCTURAL CYCLE
2018–2019: Early Hardware Milestones
└─ Shadow Hand solves Rubik's Cube via domain randomization (Dactyl project).
2020: First Disbandment
└─ Team disbanded. Co-founder Wojciech Zaremba cites lack of robot training data.
Focus shifts exclusively to large-scale language model pre-training.
May 2024: The Standalone Revival
└─ OpenAI robotics team rebuilt; heavy hiring spree across hardware and ML.
Commercial alliance forged with humanoid startup Figure AI.
November 2024: Hardware Expansion
└─ Caitlin Kalinowski hired from Meta (Orion AR lead) to head hardware engineering.
February 2025: Alliance Fracture
└─ Figure AI terminates partnership to build its own proprietary "Helix" robot AI.
March 2026: Executive Departure
└─ Kalinowski resigns over OpenAI Pentagon defense contracts and lethal autonomy.
May 2026: Capital Reallocation
└─ Internal plans to spin off hardware and robotics units shelved over balance sheet rules.
November 2026: Second Disbandment
└─ Standalone team dissolved; researchers folded into Chen (Research) & Brockman (Applied).
The 2018–2020 Phase: Dactyl and the Text Pivot
OpenAI’s early robotics efforts achieved global recognition in October 2019, when researchers used an anthropomorphic Shadow Dexterous Hand, driven by reinforcement learning within a massively randomized simulation (the Dactyl project), to solve a Rubik’s cube one-handed. The achievement relied heavily on domain randomization—varying simulated physics, friction, mass, and visual lighting across thousands of concurrent simulations so the neural network could transfer directly to physical hardware without real-world fine-tuning.
By late 2020, however, internal enthusiasm stalled. OpenAI disbanded its original robotics team, dispersing its researchers across emergent language and vision teams.
OpenAI co-founder Wojciech Zaremba addressed the closure on a machine learning podcast, acknowledging that data scarcity had created a progress wall:
"There are actually plenty of domains that are very, very rich with data. And ultimately that was holding us back in terms of robotics."
Internet text, code repositories, and image datasets were accessible at internet scale; physical robotic interaction data was virtually nonexistent, expensive to collect, and legally ambiguous. The decision to dissolve the robotics team in 2020 allowed OpenAI to reallocate computing clusters toward scaling transformer models, a redirection that culminated directly in the release of ChatGPT in November 2022.
The 2024 Revival and the Figure AI Divorce
By early 2024, the commercial and technical landscape had altered dramatically. Venture capital flooded into the humanoid robotics space, driven by advances in Vision-Language-Action (VLA) models and generative video architectures.
Fearing it had surrendered physical artificial intelligence to rivals like Tesla, Google DeepMind, and emergent hardware startups, OpenAI systematically rebuilt a dedicated robotics team starting in May 2024. The company hired dozens of roboticists, simulation experts, and mechanical engineers, building out physical teleoperation laboratories in San Francisco.
OpenAI concurrently entered a partnership with Figure AI, the humanoid developer founded by Brett Adcock. OpenAI backed Figure financially, and the two companies published demonstration videos showing the Figure 01 humanoid engaging in natural speech, semantic scene parsing, and physical item manipulation powered by OpenAI visual-reasoning models.
The alliance deteriorated within a year. In February 2025, Figure officially ended its commercial partnership with OpenAI.
Adcock and Figure’s leadership determined that depending on a third-party foundation model provider left them vulnerable to API latencies, high token costs, and a lack of granular control over low-level motor tokens. Figure announced it would develop its own proprietary embodied AI stack—an end-to-end foundation model designated Helix—optimized explicitly for sub-millisecond robotic motor control.
The loss of Figure forced OpenAI back to the drawing board. Deprived of its primary external humanoid platform, the company was forced to either build its own full-stack mechanical hardware or double down on proving that general-purpose frontier models could control arbitrary third-party machines out of the box.
Executive Turbulence: Hardware Departures and the Defense Rift
The path leading to the November restructuring was further complicated by high-profile executive turnover and internal ethical controversies surrounding the company’s expansion into defense hardware.
In November 2024, OpenAI recruited Caitlin Kalinowski to lead its robotics and consumer hardware divisions. A respected Silicon Valley hardware executive, Kalinowski had previously spent nearly a decade at Apple leading mechanical design for the MacBook Air and Mac Pro lines, before serving as Meta’s head of VR/AR hardware, where she oversaw the engineering of the Oculus Quest series and Meta's Orion augmented reality glasses prototype.
At OpenAI, Kalinowski was tasked with transforming the software-heavy company into an organization capable of designing, sourcing, and mass-producing physical hardware. She quietly scaled a San Francisco facility employing more than 100 teleoperators and data-collection technicians who guided robotic limbs through thousands of daily household and industrial chores.
CAITLIN KALINOWSKI RESIGNATION POST
"I care deeply about the Robotics team and the work we built together...
AI has an important role in national security. But surveillance of
Americans without judicial oversight and lethal autonomy without human
authorization are lines that deserved more deliberation than they got.
This was about principle, not people."
— Caitlin Kalinowski, March 7, 2026
That hardware buildout fractured abruptly in March 2026. Following negotiations between the U.S. Department of Defense and foundation model developers, rival firm Anthropic refused Pentagon contract terms, citing corporate ethical policies prohibiting the deployment of its models for domestic mass surveillance or fully autonomous lethal weapon systems. Washington moved to limit federal use of Anthropic’s technology, creating an opening that OpenAI quickly filled by finalizing an agreement to deploy its frontier models across classified defense networks.
The deal provoked internal backlash. On March 7, 2026, Kalinowski posted her public resignation to social media:
"I care deeply about the Robotics team and the work we built together. This wasn't an easy call. AI has an important role in national security. But surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got. This was about principle, not people."
Her departure left a leadership vacuum in OpenAI's hardware engineering group. Without an executive of Kalinowski’s operational caliber to oversee manufacturing supply chains, mechanical prototyping, and sensor integration, the internal friction of maintaining a standalone hardware unit intensified.
Rather than hiring a direct replacement to run an isolated hardware vertical, OpenAI’s executive committee began debating whether the company should be building physical machines at all.
The Financial Realities: Compute Prioritization and Pre-IPO Austerity
Behind the architectural theories and executive departures lies a stark financial reality: OpenAI is executing an aggressive capital-efficiency push ahead of an expected initial public offering in early 2027.
In the third quarter of 2026, OpenAI recorded operating losses of $10.9 billion, driven primarily by the astronomical compute expenditures required to train next-generation frontier architectures and run inference across millions of commercial enterprise accounts. The company is actively executing another major private funding round to bridge capital requirements before public markets open, and prospective institutional investors have demanded strict cost discipline and operational focus.
FINANCIAL AND STRATEGIC PRESSURES (Q3–Q4 2026)
• Operating Losses: $10.9 Billion in Q3 2026
• Capital Mandate: CEO Sam Altman orders teams to eliminate internal "side quests"
• Spin-Off Stalled: May 2026 effort to spin off hardware blocked by balance-sheet rules
• IPO Horizon: Corporate restructuring aimed at early 2027 public market debut
• Resource Shift: Thousands of GPUs reallocated from robotics simulation to GPT-6
CEO Sam Altman has repeatedly instructed senior staff over recent months to abandon what he characterized as internal "side quests". A dedicated robotics hardware effort is arguably the most capital-intensive side quest in tech: it demands custom fabrication, precision actuator testing facilities, mechanical maintenance, teleoperation labor fleets, and thousands of dedicated graphics processing units (GPUs) running high-dimensional physics simulations.
In May 2026, OpenAI management explored an alternative corporate path: spinning off its consumer hardware and robotics divisions into an independent, venture-backed startup entity. The blueprint resembled OpenAI’s structure with early vehicle partners, allowing the spinoff to raise non-dilutive outside venture capital while maintaining exclusive API agreements with OpenAI.
The plan was shelved when corporate auditors and legal advisors determined that standard accounting consolidation rules would still require OpenAI to report the spin-off’s massive mechanical capital expenditures and operational liabilities directly on its consolidated pre-IPO balance sheet.
Faced with the choice between carrying an expensive, slow-moving physical manufacturing division into public disclosures or folding the workforce into high-margin software groups, Altman and Brockman chose the software path.
Disbanding the standalone unit freed thousands of clustered GPUs previously earmarked for robotics reinforcement learning, instantly redirecting compute capacity toward core commercial API acceleration and model inference.
Architectural Deep Dive: How General Models Actuate Physical Systems
OpenAI’s technical rationale hinges on fundamentally transforming how neural networks interface with physical hardware. To understand why the company believes a standalone robotics team is unnecessary, one must examine how general-purpose frontier systems are shifting from pure natural language and vision processing into direct physical control.
TRADITIONAL VS. DIRECT GENERAL-PURPOSE ROBOT CONTROL
Traditional Modular Robotics Pipeline:
[Sensors/RGBD] ──> [Perception Model] ──> [Task Planner] ──> [Trajectory Optimization] ──> [PID Controller] ──> [Motors]
(Requires specialized code, brittle interfaces, manually tuned control layers)
OpenAI's Direct Foundation Model Pipeline:
[Multimodal Sensory Tokens] ──> [ Unified GPT-6 Astra Core ] ──> [ Tokenized Motor Actions ] ──> [Low-Level Microcontroller]
(Perception, semantic reasoning, spatial planning, and motor policies unified in one network)
Historically, robotics software architectures relied on deeply modular, manually engineered pipelines:
- Perception Modules: Processed raw camera pixels and depth maps into segmented point clouds and labeled 3D bounding boxes.
- Semantic Task Planners: Ingested the bounding boxes and formulated logical steps (e.g., "move to coordinate X, open gripper, translate along Z-axis").
- Trajectory Generators: Computed inverse kinematics, avoiding joint singularities and physical collisions using classical motion planners like RRT (Rapidly-exploring Random Trees).
- Low-Level Motor Controllers: Relied on proportional-integral-derivative (PID) or computed torque control loops running on microcontrollers to regulate current and voltage to physical brushless motors.
This modular structure required dedicated robotics specialists at every seam to tune kinematics, calibrate sensor fusion, and write exception handling for unpredictable edge cases.
The Vision-Language-Action (VLA) Transition
Over the last two years, embodied AI research converged toward end-to-end Vision-Language-Action (VLA) models. Under this framework, continuous physical actions are tokenized into discrete values, directly mirroring text tokens:
$$\mathcal{A}_t = \{a_1, a_2, \dots, a_k\} \quad \text{where} \quad a_i \in \text{Quantized Joint Displacements}$$
A single transformer model ingests an image:
$$I_t \in \mathbb{R}^{H \times W \times C}$$
and a textual task prompt:
$$\mathcal{T}$$
It then outputs a string of predictive action tokens:
$$\mathcal{P}(\mathcal{A}_t \mid I_t, \mathcal{T})$$
These tokens represent 7-degree-of-freedom (7-DoF) end-effector poses or individual joint angle targets.
OpenAI’s breakthrough with GPT-6 Astra was demonstrating that an ultra-scale multimodal model, trained on trillions of tokens of web-scale video, code, text, diagrams, and simulated interactions, already develops rich internal geometric representations. Astra does not view an image of a table merely as an arrangement of pixels; its attention heads inherently encode spatial depth, support surfaces, center-of-mass distributions, and structural affordances.
Hierarchical Decoupling: Action Chunking via Diffusion
The central engineering barrier remains latency. A massive frontier model cannot run inference at the 50 Hertz required to balance a bipedal robot or adjust a mechanical grip as an egg slips between rubber fingertips.
To bridge this operational divide without an isolated robotics software team, OpenAI’s Applied organization is standardizing a two-tier Hierarchical Control Architecture:
HIERARCHICAL CONTROL STACK
┌──────────────────────────────────────────────┐
│ High-Level Reasoning: GPT-6 Astra Core │
│ • Operates at 0.5 to 1.0 Hz (Cloud/Data Ctr)│
│ • Analyzes complex semantic instructions │
│ • Emits long-horizon Action Chunks │
└──────────────────────┬───────────────────────┘
│
▼ Action Latent Trajectories
┌──────────────────────────────────────────────┐
│ Mid-Level Policy: Local Diffusion Policy │
│ • Operates at 50 Hz (Onboard Edge GPU) │
│ • Unrolls continuous trajectories via ACT │
│ • Corrects for slight kinematic deviations │
└──────────────────────┬───────────────────────┘
│
▼ Joint Position Commands
┌──────────────────────────────────────────────┐
│ Low-Level Execution: Microcontroller Loops │
│ • Operates at 1,000 Hz (Actuator Level) │
│ • Torque control, current regulation │
│ • Hardware impedance & force safety stops │
└──────────────────────────────────────────────┘
- The Slow Asynchronous High-Level Planner (0.5 – 1.0 Hz): The centralized Astra foundation model runs on cloud infrastructure or high-end local edge servers. It evaluates the scene, plans sub-goals, and outputs long-horizon "action chunks"—dense latent vectors describing intended trajectory envelopes over 2 to 5-second intervals.
- The Fast Mid-Level Policy (50 Hz): Running directly on low-power onboard hardware (such as an embedded Nvidia Jetson Thor or custom ASIC), a lightweight, locally deployed diffusion policy or Action Chunking Transformer (ACT) receives Astra’s latent vectors. It decodes them into smooth continuous motor trajectories, dynamically reacting to local vision sensors.
- Low-Level Reflex Loops (1,000 Hz): Dedicated microcontrollers execute field-oriented control (FOC) algorithms to regulate motor currents, preventing mechanical over-torque and maintaining dynamic balance.
By adopting this hierarchical model, OpenAI no longer requires an internal cohort of engineers to construct specialized neural network backbones for robotics. The core research team improves the reasoning engine at the top; third-party hardware manufacturers or the Applied engineering group under Brockman handle the local mechanical adapters at the bottom.
Academic and Industrial Reactions: A Contested Hypothesis
OpenAI’s structural pivot has drawn sharp commentary from roboticists and AI researchers worldwide. The split in opinion mirrors a fundamental debate: can the intricacies of physical interaction be solved purely through scale, or does physical hardware impose unique constraints that software scaling cannot resolve?
EXPERT PERSPECTIVES ON OPENAI'S ROBOTICS RETREAT
Guanya Shi (Carnegie Mellon University Robotics Institute):
"OpenAI's posture indicates they view robotics as no longer a special
problem. In their view, physical actuation is merely a downstream
modality that will be solved naturally as foundation models scale."
Russ Tedrake (MIT / Walden Robotics):
"A simulated manipulation benchmark tells you almost nothing about
real-world reliability. Specialized startups that operate physical
fleets every day maintain a massive deployment advantage over cloud
models dealing with latency and real-world friction."
Pete Florence (Cofounder, Generalist):
"It remains completely unclear that any existing general model will
dominate robotics. Physical dynamics, contact mechanics, and microsecond
tactile feedback require fundamentally different training signals."
Guanya Shi, an assistant professor at Carnegie Mellon University’s Robotics Institute, observed that the move highlights OpenAI’s unwavering institutional faith in the scaling hypothesis:
"OpenAI’s posture indicates they view robotics as no longer a special problem. In their view, physical actuation is merely a downstream modality that will be solved naturally as foundation models scale."
Shi points out that if this premise is correct, maintaining a dedicated robotics team is inefficient. Every engineering hour spent tuning custom kinematics on a physical bench is an hour diverted from improving the core pre-training architectures that lift all modalities simultaneously.
Conversely, Russ Tedrake, professor of electrical engineering and computer science at MIT and founder of Walden Robotics, argued that OpenAI is underestimating the unforgiving physical realities of the mechanical domain:
"A simulated manipulation benchmark tells you almost nothing about real-world reliability. Specialized startups that operate physical fleets every day maintain a massive deployment advantage over cloud models dealing with latency and real-world friction."
Tedrake emphasizes that simulation environments, no matter how photorealistic or dynamically complex, cannot accurately capture the chaotic variables encountered on physical factory floors or domestic living spaces: dust accumulation on lenses, micro-slippages in gear trains, temperature-induced sensor drift, and compliant physical interactions.
Pete Florence, cofounder of robotics AI startup Generalist and a veteran of Google's early robotics efforts, expressed skepticism regarding the universal viability of standard foundation model backbones:
"It remains completely unclear that any existing general model will dominate robotics. Physical dynamics, contact mechanics, and microsecond tactile feedback require fundamentally different training signals than predicting internet text or synthetic video frames."*
Within industry circles, several robotics founders welcomed OpenAI’s retreat from dedicated hardware development, interpreting the decision as a validation of specialized embodied AI firms. With OpenAI abandoning proprietary hardware pipelines to act as an infrastructure layer, independent robotics developers face less immediate risk of being directly undercut by an integrated, vertically closed OpenAI machine.
The Humanoid Competitive Landscape: Diverging Strategies
The dissolution of the dedicated OpenAI robotics team sharpens the strategic divergence across the robotics sector. As tech titans and venture-backed startups race to commercialize general-purpose humanoid labor, three contrasting philosophical camps have emerged:
THE THREE COMPETING EMBODIED AI STRATEGIES
Strategy 1: Full-Stack Vertical Integration
┌────────────────────────┐ • Companies: Tesla (Optimus), Figure AI (Helix)
│ Proprietary Hardware │ • Philosophy: The brain and body must be co-designed.
│ ▲ │ End-to-end models trained specifically on fleet
│ ▼ │ teleoperation data. Hardware and inference silicon
│ Proprietary Robot AI │ built exclusively for each other.
└────────────────────────┘
Strategy 2: Hardware-Centric Agnosticism
┌────────────────────────┐ • Companies: Boston Dynamics (Atlas), Agility (Digit)
│ Superior Hardware & RL │ • Philosophy: Mechanical perfection, dynamic stability,
│ ▲ │ and deterministic safety first. High-level AI
│ ▼ │ brains are modular, swappable components integrated
│ Swappable AI Brains │ via enterprise APIs.
└────────────────────────┘
Strategy 3: Universal Foundation Substrate (OpenAI Approach)
┌────────────────────────┐ • Companies: OpenAI, Google DeepMind
│ Unified Frontier Model │ • Philosophy: Mechanical hardware is a commoditized
│ ▲ │ edge device. General-purpose cloud/edge intelligence
│ ▼ │ will natively operate arbitrary mechanical systems
│ Any Third-Party Device │ via standardized interface layers.
└────────────────────────┘
1. The Full-Stack Verticalists: Tesla and Figure AI
Pioneered by Tesla’s Optimus program and aggressively adopted by Figure AI following its split from OpenAI, this philosophy argues that building an operational humanoid requires complete vertical integration.
Tesla designs its own custom structural actuators, battery packs, sensor suites, and onboard inference microchips (Dojo and System-on-Chip hardware). The Optimus neural networks are trained exclusively on end-to-end video-to-action data harvested from thousands of hours of proprietary human teleoperation and deployed vehicle fleet data.
Figure AI’s Helix model follows an identical logic: by controlling both the actuation physics and the model weights, Figure can optimize inference execution directly at the actuator level, bypassing the high token costs and latency bottlenecks inherent in generalized cloud systems.
2. The Dynamic Specialists: Boston Dynamics and Agility Robotics
Companies with deep mechatronic pedigrees, such as Boston Dynamics (backed by Hyundai) and Agility Robotics, prioritize physical reliability, kinematic durability, and industrial safety certifications above all else.
The electric Boston Dynamics Atlas relies on world-class reinforcement learning control policies designed to ensure real-time dynamic balance, omnidirectional walking, and high-speed manipulation in harsh industrial environments. For these manufacturers, high-level multimodal reasoning is a modular component: their robots can ingest task commands from an OpenAI API or a Google Gemini model, but the low-level stabilization and motion planning remain locked behind deterministic, battle-tested kinematic software guards.
3. The Pure Platform Substrate: OpenAI’s Ecosystem Play
OpenAI’s strategy abandons the burden of mechanical manufacturing entirely, seeking to establish its models as the ubiquitous software substrate across the robotics industry. Rather than bearing the manufacturing costs, warranty liabilities, and physical supply chain risks of custom hardware, OpenAI intends to capture value by billing robotic hardware companies on a per-token or monthly compute licensing basis.
This horizontal playbook mirrors the early personal computing and smartphone eras:
- Tesla and Figure are pursuing the Apple model, tightly coupling proprietary silicon, custom hardware, and exclusive software into a closed, highly optimized ecosystem.
- OpenAI is attempting the Microsoft Windows or Google Android model, betting that open-market mechanical commoditization will eventually drive hardware costs down, leaving the provider of the universal foundation model brain in control of the most profitable layer of the value chain.
Technical Bottlenecks: Moravec’s Paradox at Hyperscale
OpenAI’s bet that general models will absorb robotics relies on overcoming Moravec’s Paradox—the longstanding observation in artificial intelligence that high-level reasoning tasks (playing chess, passing legal exams, synthesizing code) require minimal computational effort, whereas low-level sensorimotor tasks (holding an apple without crushing it, sweeping uneven ground, navigating a crowded doorway) demand immense, finely tuned computational resources.
In shifting robotics into general research, OpenAI researchers must resolve three profound algorithmic and physical bottlenecks:
1. The Discrepancy of Data Geometry
Language and code models scale reliably because the underlying data is symbolic, discrete, and abundant. High-dimensional physical interaction data, by contrast, is continuous, multimodal, and non-stationary.
When a robot interacts with physical matter, it alters its environment dynamically. A slight error in torque application propagates forward unpredictably: a glass slips, shattering on the floor, instantaneously invalidating the model's forward trajectory plan. Internet text datasets contain no analog to this continuous physical divergence.
While OpenAI’s Sora and related video models can generate plausible subsequent frames, they frequently hallucinate impossible physics—objects morphing through one another, momentum vanishing, weights defying gravity. In a video output, an impossible frame is merely a visual artifact; on a physical assembly line, an impossible action command strips a gearbox or injures a human collaborator.
2. High-Frequency Tactile and Proprioceptive Integration
Human manipulation relies profoundly on subconscious touch. Humans adjust their grip forces within milliseconds of detecting microscopic shear vibrations along the skin's ridges, long before visual information can travel through the optic nerve to the cerebral cortex.
Current foundation model architectures are overwhelmingly vision-centric. They process camera feeds at frame rates between 10 and 30 frames per second.
To operate with human dexterity, models must natively ingest high-frequency, non-visual telemetry:
- Multi-axis tactile array voltages (distributed tactile skins)
- Back-EMF (electromotive force) signals from brushless DC motors
- Joint-level strain gauge readings (6-axis force-torque load cells)
Integrating continuous, high-frequency, non-visual sensor streams into a unified transformer attention context alongside visual and language tokens remains an open research frontier. The current transformer architecture struggles with the context length expansion that occurs when hundreds of sensor readings stream in at 1 Kilohertz.
3. Safety, Determinism, and Catastrophic Actions
In pure software environments, foundation model hallucinations produce incorrect code, misleading text, or synthetic imagery errors—mistakes that can be caught by unit tests or human review. In the physical realm, model hallucinations yield catastrophic structural failures:
- Driving an industrial manipulator directly into a metal CNC enclosure at maximum velocity.
- Misjudging human proximity and crushing a worker’s limb against a warehouse racking upright.
- Executing an unbalanced bipedal step on an incline, causing an 80-kilogram humanoid to fall onto bystanders.
Industrial safety regulators—including the Occupational Safety and Health Administration (OSHA) in the United States and equivalent bodies internationally—require deterministic, mathematically provable safety stops (Performance Level d/e under ISO 13849).
Because deep neural networks and generative foundation models operate as probabilistic black boxes, certifying a direct-action foundation model for ungated, close-proximity human collaboration remains legally and practically impossible without traditional deterministic software overrides.
Looking Ahead: The 2027 In-House Humanoid and Unresolved Questions
Despite the formal dissolution of the dedicated OpenAI robotics team, the company has explicitly not abandoned its hardware ambitions.
An OpenAI spokesperson confirmed that the company continues work on an internal humanoid robotics platform, with CEO Sam Altman maintaining a personal target for a first public hardware demonstration in 2027.
OPENAI'S PROJECTED ROADMAP TO 2027
2026 Q4 (Immediate):
• Complete organizational migration of robotics researchers under Mark Chen.
• Transition applied hardware engineers into Greg Brockman's Applied wing.
• Accelerate inference optimization to compress Astra's 18-second latency.
2027 H1 (Pre-IPO Window):
• File consolidated financial disclosures for prospective early 2027 IPO.
• Release unified Hardware API specification for third-party robotics manufacturers.
• Establish standardized low-latency Edge-to-Cloud hierarchical runtime.
2027 H2 (Public Hardware Milestone):
• Target window for Sam Altman's public demonstration of OpenAI's internal humanoid.
• Validation test: Does direct foundation intelligence translate into reliable physical utility?
The reorganization fundamentally redefines the nature of that 2027 milestone. Rather than unveiling a proprietary, fully custom mechanical platform developed in an insular robotics laboratory, any hardware OpenAI demonstrates in 2027 will serve primarily as a physical proof-of-concept for its centralized multimodal foundation models.
The demonstration will be structured to prove to industrial manufacturers, logistics operators, and consumer hardware developers that they no longer need to spend billions constructing bespoke robotics AI departments. OpenAI’s message will be clear: buy standard actuators and metal frames off the shelf, install an OpenAI edge runtime, and connect your machines to our centralized models.
As OpenAI moves toward its anticipated public offering, several pivotal milestones will determine whether this strategic pivot succeeds:
- Compression of Action Latency: Can Mark Chen’s core research teams distill multi-billion-parameter foundation models to run local action loops under 50 milliseconds, eliminating the crippling 18-second pauses exposed in the Astra Nvidia benchmarks?
- Inference Unit Economics: Will OpenAI achieve the architectural efficiency required to reduce the cost of robotic operations from $36.40 per attempt down to micro-cents per action cycle?
- Third-Party Hardware Partnerships: Following the collapse of the Figure AI relationship, can Brockman’s Applied team recruit major industrial robotics players—such as ABB, Fanuc, KUKA, or Hyundai/Boston Dynamics—to build official integration stacks for OpenAI’s foundation models?
- Regulatory and Safety Certification: How will OpenAI reconcile the probabilistic hallucinations inherent in generative models with the strict, deterministic physical safety standards mandated across global manufacturing and consumer sectors?
The restructuring reflects OpenAI's core conviction that artificial general intelligence is a unified computational challenge rather than a collection of separate domains. By dismantling its dedicated team, the company is declaring that the physical world will bend to the same scaling curves that transformed natural language and visual generation.
Whether that scaling thesis holds true when confronted with real friction, inertia, sensor noise, and mechanical wear on physical factory floors will determine if OpenAI has forged the operating system for future machines—or surrendered the physical world to specialized robotics competitors.
Reference:
- https://aiweekly.co/alerts/openai-folds-robotics-team-into-main-research-and-applied-orgs
- https://fortune.com/2026/11/13/openai-disbands-robotics-team-as-its-general-models-start-running-robots/
- https://www.benzinga.com/markets/private-markets/26/11/63266906/openai-dismantles-standalone-robotics-team-as-it-bets-on-bigger-ai-models
- https://mlq.ai/news/openai-disbands-its-standalone-robotics-team-as-its-general-models-start-running-robots-fortune-reports/
- https://dealroom.co/news/172712-openai-disbands-robotics-team-bets-general-models-can-run-robots/
- https://startupfortune.com/openai-disbands-its-stand-alone-robotics-team-and-moves-researchers-into-core-groups/
- https://timesofindia.indiatimes.com/technology/tech-news/openai-shuts-down-robot-unit-second-time-in-six-years-but-wants-everyone-to-know-that-the-company-is-committed-to-/articleshow/135846511.cms
- https://www.youtube.com/shorts/XegGEf7N_vM
- https://superpowerdaily.com/posts/openai-disbands-its-robotics-team-betting-general-ai-models-can-run-robots
- https://www.businessinsider.com/caitlin-kalinowski-quits-openai-robotics-head-pentagon-deal-sam-altman-2026-3
- https://fortune.com/2026/03/07/openai-robotics-leader-caitlin-kalinowski-resignation-pentagon-surveillance-autonomous-weapons-anthropic/
- https://en.wikipedia.org/wiki/Caitlin_Kalinowski
- https://www.caitlinkalinowski.com/