The Data Scarcity Problem
Why Humanoid Robots Can’t Just ‘Scrape the Internet’
In the race to build the ultimate general-purpose humanoid robot, hardware is no longer the primary bottleneck. Actuators are becoming cheaper, batteries are denser, and edge computing power is more than sufficient. The true hurdle facing the robotics industry in 2026 is a software problem—specifically, a massive deficit of training data.
While Large Language Models (LLMs) like ChatGPT achieved their seemingly magical capabilities by ingesting trillions of tokens scraped from the internet, physical AI cannot follow the same playbook. A robot cannot learn the exact torque required to grasp a fragile egg, or the subtle balance adjustments needed to walk on uneven gravel, by reading Wikipedia articles.
This is the data scarcity problem. Training a physical AI requires embodied data: synchronized streams of video, joint positions, motor torques, and tactile feedback. Because this data does not exist naturally on the internet, robotics companies must manufacture it from scratch. This article explores why training physical AI is exponentially harder than training text models, and how the industry is using teleoperation, simulation, and synthetic data to bridge the gap.
The LLM vs. Physical AI Scale Disconnect
To understand the magnitude of the data scarcity problem, one must look at the scale of existing datasets.
Modern LLMs are trained on the equivalent of roughly 100,000 years of human reading material. Text is abundant, perfectly structured, and easy to scrape. In stark contrast, the largest open-source robotics dataset in the world—Google DeepMind’s Open X-Embodiment dataset—contains just over one million real robot trajectories across 22 different robot types. While impressive, this dataset is orders of magnitude smaller than the text datasets used to train even the earliest iterations of GPT.
LLMs excel at syntax, semantics, and logic, but they lack “embodied affordances.” As robotics researchers frequently point out, an LLM can write a flawless essay about how to use a hammer, but it does not know what a handle actually feels like, nor does it understand the physical constraints of swinging one.
The following table illustrates the stark differences between training text-based AI and physical AI.
|
Metric
|
Large Language Models (Text)
|
Physical AI (Robotics)
|
|
Primary Data Source
|
Internet scraping (CommonCrawl, Reddit, books)
|
Teleoperation, simulation, proprietary collection
|
|
Data Structure
|
1D text tokens
|
Multi-modal (video, torque, depth, tactile)
|
|
Dataset Scale
|
Trillions of tokens (massive abundance)
|
Millions of trajectories (extreme scarcity)
|
|
Cost of Acquisition
|
Near zero (automated web scraping)
|
Highly expensive (human labor, physical hardware)
|
|
Failure Consequence
|
Hallucination or bad text
|
Physical damage to robot, property, or humans
|
To solve this scarcity, the humanoid robotics industry has coalesced around three primary data generation strategies: teleoperation, simulation, and world model pre-training.
Strategy 1: The Teleoperation Sweatshop
If you cannot scrape the data, you must create it manually. Teleoperation is currently the gold standard for collecting high-quality, real-world robot training data.
In a teleoperation pipeline, a human operator wears a virtual reality headset and motion-capture gloves (or a full exoskeleton suit) to remote-control a humanoid robot. As the human operator performs a task—such as folding laundry, picking up a box, or operating a coffee machine—the robot’s sensors record every micro-adjustment in joint angle, motor torque, and visual perspective. This data is then fed into the neural network, allowing the AI to learn by mimicking human demonstrations.
This approach yields excellent results, but it is incredibly expensive and difficult to scale. In 2026, companies like Tesla are paying “Data Collection Operators” up to $48 per hour to wear motion-capture suits and perform repetitive tasks. Other companies, such as 1X Technologies and Prosper Robotics, have begun paying gig workers to collect teleoperation data from their own homes, effectively crowdsourcing the physical labor required to train the AI.
Sanctuary AI, the Canadian company behind the Phoenix humanoid, has built its entire AI development strategy around teleoperation. Now in its eighth generation, Phoenix improves with each iteration because Sanctuary has accumulated years of high-quality teleoperation data. The company’s Genesis System is designed to mimic human learning processes, using teleoperation demonstrations as the foundational curriculum for its AI.
Despite these efforts, teleoperation alone cannot generate the billions of data points required for true general-purpose autonomy. It is simply too slow and too costly. A single hour of teleoperation data can cost anywhere from $25 to $200 depending on the complexity of the task and the skill of the operator. To match the scale of LLM training data, the industry would need millions of operator-hours—an economically prohibitive proposition.
The Teleoperation Data Pipeline
The typical teleoperation data pipeline follows a structured sequence. First, a human operator performs a task while wearing motion-capture equipment. The robot’s sensors simultaneously record joint positions, motor currents, camera feeds, and (where available) tactile pressure. This raw data is then cleaned, labeled, and formatted into training episodes. Finally, the episodes are used to train or fine-tune a Vision-Language-Action (VLA) model through imitation learning.
The quality of this data is paramount. As Claru AI noted in a recent analysis, data quality consistently beats data volume in robotics training. A single high-fidelity demonstration of a complex manipulation task can be worth more than a thousand sloppy ones.
Strategy 2: Simulation and the “Sim-to-Real” Gap
To achieve the scale of an LLM, robots must be trained in the digital world. Simulation allows developers to run thousands of virtual robots simultaneously, running millions of trial-and-error experiments in accelerated time without the risk of breaking expensive hardware.
NVIDIA has positioned itself as the dominant force in this arena. The company’s Isaac Sim platform serves as the foundational virtual environment for testing and training AI-based robots. At the NVIDIA GTC 2026 conference, the company unveiled the Physical AI Data Factory Blueprint—an open reference architecture designed to unify data curation, augmentation, and evaluation.
A critical component of this blueprint is the generation of synthetic data. Using NVIDIA’s Cosmos 3 world foundation model and the Isaac GR00T-Dreams system, developers can generate massive amounts of synthetic trajectory data from just a single image. For example, if a robot needs to learn how to pick up a mug, the simulation can automatically generate thousands of variations of that mug—different colors, shapes, lighting conditions, and background environments.
This technique, known as “domain randomization,” forces the AI model to learn the invariant features of a task rather than memorizing a specific environment. However, simulation is not a silver bullet. The physics engine in a simulation, no matter how advanced, never perfectly matches the messy reality of the physical world. Friction, sensor noise, and unexpected material deformation create a “sim-to-real gap.” If a model is trained exclusively in simulation, it will almost certainly fail when deployed on physical hardware.
To combat this, researchers employ a combination of techniques. Domain randomization varies the visual and physical properties of the simulated environment—textures, lighting, friction coefficients, object masses—across thousands of training runs. This forces the model to learn robust, generalizable behaviors rather than memorizing a single environment. Reinforcement learning co-training, a technique validated in a March 2026 paper on VLA models, combines simulated and real-world data during training, yielding stronger generalization to unseen task variations.
Allen AI’s MolmoBot project, released in March 2026, demonstrated that it is now possible to train robot manipulation policies entirely in simulation and deploy them with zero real-world fine-tuning. This represents a significant milestone, though the approach currently works best for structured manipulation tasks rather than the open-ended behaviors required for general-purpose humanoids.
Strategy 3: World Models and Video Pre-Training
To bridge the gap between expensive teleoperation and imperfect simulation, companies are turning to “World Models” trained on internet video.
While robots cannot learn physical torques from text, they can learn an intuitive understanding of physics from video. 1X Technologies, the maker of the NEO home robot, has pioneered this approach. The company’s proprietary World Model acts as a data-driven simulator built with a grounded understanding of physics.
According to 1X, their model is pre-trained on over one million hours of general internet video showing humans performing everyday tasks. This is supplemented by 900 hours of human first-person perspective video and hundreds of hours of unfiltered, random robot play data. By watching millions of hours of humans interacting with the world, the AI develops a baseline understanding of object permanence, gravity, and spatial relationships. Once this foundation is established, the model requires significantly less teleoperation data to master a specific physical task.
Company Data Strategies Compared
The following table compares how leading humanoid companies are approaching the data scarcity problem, revealing the diversity of strategies and the emerging consensus around hybrid approaches.
|
Company
|
Primary Strategy
|
Simulation Platform
|
Teleoperation
|
Video Pre-Training
|
Key Differentiator
|
|
1X Technologies
|
World Model + Teleop
|
NVIDIA Isaac Sim
|
Gig worker crowdsourcing
|
1M+ hours internet video
|
Video-first foundation; lowest cost per data point
|
|
Tesla
|
Teleop + Sim
|
Internal (Dojo)
|
$25–$48/hr operators in motion-capture suits
|
Limited
|
Massive capital to brute-force data collection at scale
|
|
Sanctuary AI
|
Teleop-first
|
Internal
|
Genesis System (8th gen)
|
Limited
|
Highest-quality human demonstrations; years of accumulated data
|
|
Google DeepMind
|
Cross-embodiment dataset
|
Internal + Isaac Sim
|
Open X-Embodiment (21 institutions)
|
RT-2 video pre-training
|
Largest open-source dataset; cross-robot transfer learning
|
|
NVIDIA
|
Full-stack platform
|
Isaac Sim + Cosmos 3 + Newton
|
N/A (enables others)
|
Cosmos world model
|
Provides the infrastructure layer for the entire industry
|
|
Figure AI
|
Real-world deployment
|
NVIDIA Isaac Sim
|
Active deployment at BMW
|
Helix VLA model
|
11 months of real production data from BMW Spartanburg
|
|
Allen AI
|
Pure simulation
|
MolmoSpaces
|
None
|
MolmoBot VLA
|
Zero real-world data required for structured tasks
|
The table reveals a clear trend: no single strategy is sufficient. The most advanced companies are converging on hybrid pipelines that combine all three approaches—video pre-training for foundational physics understanding, simulation for scale, and teleoperation for real-world refinement.
The Emerging Data Flywheel
The most promising development in 2026 is the emergence of the “data flywheel” concept. Companies like Figure AI, which has deployed robots on the BMW assembly line for 11 months, are generating real-world training data as a byproduct of commercial operations. Every shift the robot works produces hours of sensor data that can be fed back into the training pipeline. As the robot improves, it handles more tasks, generating even more data.
This flywheel effect mirrors the dynamic that propelled LLMs: the more users interacted with ChatGPT, the more data OpenAI collected, which made the model better, which attracted more users. The humanoid companies that achieve commercial deployment first will gain an insurmountable data advantage over competitors still confined to the lab.
Conclusion
The humanoid robot revolution is currently gated by the data scarcity problem. The hardware is ready, but the physical AI brains are still starving for embodied experience.
The companies that win the humanoid race will not be those with the best motors or the sleekest chassis. The winners will be the companies that build the most efficient “data engines”—seamless pipelines that combine internet video pre-training for baseline physics, synthetic data generation in NVIDIA Isaac for scale, and high-fidelity teleoperation for real-world refinement. Until that pipeline is perfected, humanoids will remain impressive prototypes rather than ubiquitous autonomous workers. The data scarcity problem is not merely a technical challenge—it is the defining strategic battleground of the humanoid industry in 2026.