One useful way to think about humanoid robotics today is that the field may still be at an earlier stage than large language models.

For LLMs, we’ve already found a fairly reliable recipe for scaling. Larger models, more data, and more compute tend to produce better systems. The exact relationship is not perfect, but it is predictable enough that companies are spending billions of dollars with some confidence that capability will improve.

Robotics is not there yet. We’re still trying to understand what exactly should scale and how the different pieces should fit together. Because model size is not the main bottleneck. The actual challenge is generalization.

A robot can often learn a task in a controlled environment. It is much harder to make that behavior work reliably in a new environment with different objects and even slightly different conditions.

Humans handle this kind of variation easily because we carry a broad understanding of the physical world with us. We know how objects generally behave and can adapt that knowledge to situations we have never seen before.

Robotic systems need something similar. This is one reason why foundation models are becoming increasingly important in robotics. Instead of training a separate system for every individual task, the goal is to build a more general world model that can transfer what it has learned across tasks and environments.

The approach has clearly parallels with language models. Large language models became much more useful once they moved away from narrow task-specific systems and toward general models trained on broad datasets. Those models could then be adapted to many downstream tasks.

Robotics is moving in the same direction, but physical interaction creates a unique challenge. There is far less robotics data available. Language models could be trained on huge amounts of text that already existed. Robotics data is much more scarce and harder to collect because it usually requires a physical system interacting with the world.

As more robots are put into real environments, they generate more useful interaction data. That data can improve the next generation of models. Better models can then support wider deployment.

If that loop becomes strong enough, progress could accelerate.

This is also where reinforcement learning becomes important.

Demonstration data can show a robot how to perform a task. Reinforcement learning can help the system improve through repeated interaction.

Robotics requires very high reliability.

A model that works most of the time may be enough for a research demonstration. Commercial systems often need to work almost every time.

Closing that gap is difficult and likely requires a large amount of experience.

Another important change is that robotics no longer has to learn everything from physical interaction alone.

Modern vision-language models already contain a broad understanding of objects and basic relationships in the world. Robotics models can use some of that knowledge rather than learning every concept from scratch.

This is the idea behind vision-language-action models.

They combine perception and language with control. In principle, this allows the model to understand a task at a higher level before translating that understanding into actions.

But we still do not know how much robotics data is required, how much of the learning can come from internet-scale models, or which training methods will generalize best across different environments.

A modern robot requires substantial compute at the edge. It also depends on sensors, memory, power electronics, and other semiconductor components. If robotics eventually becomes a large computing platform, it could become a meaningful source of semiconductor demand