The next major bottleneck in robotics may not be motors, batteries or mechanical design. It could be data.
Humanoid robots need enormous amounts of information to learn how people move, manipulate objects and interact with different environments. As a result, startups are beginning to build businesses around collecting and organizing the data required to train physical AI systems.
One example is Mecka AI, which focuses on human-motion data for humanoid robots and other robotics systems. TechCrunch reported in September that the company was nearing a financing round at a valuation of approximately $500 million, following a $60 million financing announced earlier in 2026. (TechCrunch)
Why physical AI needs different data
Generative AI systems can learn from enormous quantities of text, images, audio and video.
Robots require something more complicated.
They need to understand relationships between perception and physical action.
For example, a robot must learn that reaching toward an object, grasping it and moving it somewhere else requires coordinated movements involving the entire body.
The same task can become significantly more difficult when the object is heavier, slippery, differently shaped or located in a new environment.
Humans become part of the training pipeline
Companies are therefore experimenting with systems that capture human movement.
Motion sensors, cameras and other technologies can record how people perform tasks. That information can then be processed into datasets that help train robotic systems.
The goal is to provide robots with examples of how humans solve physical problems.
From robotics hardware to a data ecosystem
This creates a new layer in the robotics industry.
The ecosystem increasingly includes:
Robot hardware
Sensors and actuators
AI models
Simulation environments
Training data
Data-collection platforms
Robot-control software
Deployment and fleet-management systems
Mecka’s funding activity illustrates how investors are beginning to finance this infrastructure alongside robot manufacturers. (TechCrunch)
The challenge of transferring data into reality
Training data alone does not guarantee that a robot will work in the real world.
Physical environments contain unpredictable variables that simulations may not capture.
A model trained on thousands of examples still needs to understand how those examples translate into new situations.
That is why real-world deployment is becoming increasingly important.
The future of robotics may therefore depend not just on building better machines, but on building a continuous pipeline of data → models → robots → real-world experience → new data.