Physical AI systems need something that doesn't exist yet: billions of hours of real-world robot interaction data captured under actual deployment conditions. While foundation models like NVIDIA's GR00T N1 have grabbed headlines, the unsexy truth is that companies building proprietary datasets are positioning themselves to dominate the robotics revolution. Investment in general-purpose robotics surged fivefold between 2022 and 2024, exceeding $1 billion annually, but the infrastructure to train these systems at scale remains fragmented and incomplete.

Why Can't Robots Just Learn From the Internet Like AI Models Do?

For decades, hardware limitations held robotics back. Sensors were too expensive, actuators too imprecise, and edge computing too costly. That ceiling has effectively vanished. What replaced it is something far more stubborn: the data problem. Language models like ChatGPT scraped the entire internet for training material. Computer vision systems relied on ImageNet, a purpose-built benchmark with millions of labeled images. Robotics has neither.

The fundamental difference is physical. Every useful data point in robotics must come from an actual robot interacting with the real world under the exact conditions it will face in deployment. You cannot scrape this from the web. A robot learning to fold a shirt in a controlled lab environment will fail spectacularly in a warehouse with variable lighting, partial occlusions, and deformable materials that behave unpredictably.

"Demand for large-scale training data and annotation services is growing fastest in the robotics and embodied AI space. There is no large, readily available corpus of pre-training data. Some researchers estimate that only a fraction of the required data exists today, meaning millions of hours of annotated egocentric, multi-sensor datasets will be needed," explained Steve Nemzer, Senior Director of Artificial Intelligence Research and Innovation at TELUS Digital.

Steve Nemzer, Senior Director, Artificial Intelligence Research & Innovation, TELUS Digital

What Makes Robotics Data So Specialized and Hard to Collect?

Robot datasets are fundamentally different from any other AI training material. They are egocentric, meaning they capture the world from the robot's perspective, and they combine multiple sensor types that must remain perfectly synchronized:

• Camera: Captures visual information about objects, surfaces, and spatial layout
• Lidar: Provides 3D distance measurements and depth perception in varying light conditions
• Radar: Detects motion and objects through occlusion and adverse weather
• Touch sensors: Record contact pressure and surface texture feedback
• Force and torque data: Measure grip pressure and resistance, critical for fine-grained manipulation tasks

For delicate tasks like plugging in a cable or peeling a label off a curved surface, force and torque feedback becomes essential. Video alone cannot teach a robot how hard to grip or when to adjust pressure. The DROID dataset, one of the largest open robot manipulation datasets, required 50 collectors across three continents just to reach its current scale. The ROVER dataset returned to the same five outdoor locations across all four seasons specifically because lighting, vegetation, and weather patterns measurably degrade robot vision systems.

How Are Companies Building the Data Infrastructure Now?

The companies that started collecting and annotating specialized robotics data years ago now hold a structural advantage. Building this infrastructure takes time: specialized capture equipment, synchronized multi-sensor annotation tooling, compliance systems, and a trained workforce familiar with domain-specific edge cases do not arrive off the shelf.

When NVIDIA released its humanoid robot foundation model GR00T N1 in March 2025, every team attempting to deploy it faced the same constraint: fine-tuning data required to make the model useful for a specific task in a specific environment on a specific machine. Foundation models give everyone the same starting point, but proprietary datasets decide who wins.

Global digital solution providers like TELUS Digital now occupy a critical position in this landscape. With more than 1 million trained annotators across five continents, these platforms deliver multi-sensor fusion annotations and compliance infrastructure that go beyond what foundation models provide.

Steps to Prepare for Physical AI Deployment

• Decide on data ownership early: Teams must choose whether to build capture infrastructure in-house or partner with a specialized provider before deployment timelines begin, as both approaches require years of preparation
• Define use-case-specific requirements: Logistics, autonomous driving, and healthcare robotics each require different sensor configurations, annotation schemas, and compliance obligations; selecting a partner evaluated against the wrong profile delivers at the wrong specification
• Plan for real-world variability: Synthetic data can fill specific gaps, but production systems must be anchored in real-world data that captures the long tail of environmental variability that simulation cannot fully reproduce

Deloitte's 2026 Tech Trends report identifies data management and cybersecurity as primary barriers to scaling physical AI beyond early industrial deployments. Teams that make infrastructure decisions early give themselves room to stay on schedule.

Is the Market Actually Moving Beyond Prototypes?

Yes. Amazon now operates more than 1 million robots across its fulfillment network, working alongside more than 750,000 employees. CEO Andy Jassy's recent shareholder letter emphasized that robotics is a key driver of productivity, efficiency, and operational improvement, and critically, Amazon views robotics as an extension of AI rather than a separate technology initiative.

Strategic partnerships between legacy industrial companies and robotics developers are accelerating commercialization. Humanoid and Schaeffler announced an agreement including actuator supply contracts and humanoid robot purchases for Schaeffler's global manufacturing network over the next five years, aiming to deliver up to 2,000 robots by 2032. China's State Grid announced plans to deploy 8,500 embodied AI systems, including 500 humanoid robots, for utility maintenance and infrastructure inspection.

Commercial activity among robot manufacturers themselves is also accelerating. RobotEra recently achieved thousand-unit shipments while reporting growth exceeding 300 percent year-over-year. Unitree has continued expanding production capacity and commercial deployments across industrial and enterprise applications.

The latest signal of confidence comes from funding. Generalist AI, a startup building embodied robotics intelligence founded by former DeepMind senior scientist Pete Florence, raised $400 million at a $2 billion valuation, with backing from Radical Ventures, 8VC, Union Square Ventures, Hanabi Capital, NVIDIA's NVentures, and Bezos Expeditions. The company's GEN-1 foundation model, released in April, demonstrates mastery of physical tasks with adaptive intelligence that allows robots to adjust and retry when conditions change, such as when objects deform or lighting shifts unexpectedly.

Morgan Stanley estimates that humanoid robotics could ultimately become a $5 trillion global market opportunity, making it one of the largest emerging technology markets over the coming decades. Approximately $26 billion has been committed through various government-supported robotics and embodied AI investment funds across China alone, supporting everything from AI model development and advanced manufacturing to robotic components and deployment infrastructure.

The robotics revolution is no longer theoretical. It is being built on a foundation of specialized data that companies are collecting and annotating right now. Those who own or control access to that data layer will shape the next decade of physical AI.