How OpenAI, Google, and Figure Are Changing the Way Robots Learn: A Brief History of the Approach

Robotics isn't just going through another round of hardware upgrades. The underlying logic of how robots learn is changing: machines are less and less often programmed by hand and more and more often trained on data, simulations, and large multimodal models. For a long time, the industry had a visible gap between ambition and results.

Engineers dreamed of science-fiction-grade robots: general-purpose helpers that could move through the ordinary world, understand people, and safely handle dozens of tasks. In practice, the market spent decades living mostly off narrower systems: robotic arms for factories, warehouse manipulators, and household devices like robot vacuums. The problem wasn't a lack of good mechanics.

Robots simply struggled with unpredictable environments. Any deviation from the script, whether a new object, a different camera angle, or a change in lighting, could quickly break their behavior. The first big shift came with trial-and-error learning.

In the 2010s, roboticists began making heavy use of reinforcement learning and simulators, where a machine can practice millions of times without risking damage to expensive hardware. One telling example is OpenAI's Dactyl, a robotic hand trained in simulation to manipulate a Rubik's Cube before the skill was transferred to the real world. The key idea was domain randomization: during training, researchers varied the environment's parameters, including friction, object mass, camera imagery, and other details.

That way, the robot learned not one ideal scenario but robust behavior across a wide range of conditions. In parallel, learning from demonstration matured as well: instead of spelling out every rule by hand, researchers began teaching robots by showing them human actions. But that still wasn't enough.

Simulation sped up training, but the gap between virtual and physical environments never went away, and real-world data remained expensive and scarce. So the next phase began when robotics picked up the techniques behind the large-model boom. Instead of building separate systems for single tasks, researchers started assembling large datasets of trajectories, images, text instructions, and actions.

Google trained its RT-1 model on 130,000 episodes collected by a fleet of 13 robots, covering more than 700 tasks. RT-2 went further: it combined robotics data with web data from vision-language models, so a robot could not only repeat a familiar motion but also draw more general conclusions from a natural-language command. The same turn shows up in projects that try to scale learning across different types of machines at once.

For the Open X-Embodiment dataset, researchers gathered data from 22 robots across 21 institutions and described hundreds of skills, so models could transfer experience between platforms instead of starting from scratch for every new manipulator. On top of this sits the idea of foundation models for the physical world: a single base intelligence that gets fine-tuned for a specific body, gripper, or work environment. Startups such as Figure, Apptronik, Covariant, and Physical Intelligence are now building not just research programs but commercial strategies around this idea.

Against this backdrop, the market has come back to life. In 2025, companies and investors put $6.1 billion into humanoid robots, four times more than the year before.

The money isn't flowing because robots can already do everything. It's flowing because there is now a more plausible path to training them further in real-world environments. Still, a general-purpose robot remains a long way off.

Machines still struggle with long sequences of actions, fine manipulation of several objects at once, working in cluttered spaces, and interacting safely with people without a pre-arranged scene. Unlike a language model, a robot can't just make a mistake in text: its errors mean a dropped object, a breakdown, downtime, or a risk to a person. That's why the industry is taking a hybrid route: more simulation, more real-world data from robot fleets, more general-purpose models, but also more constraints from safety and economics.

The main takeaway is simple: today's breakthrough in robotics has less to do with a new kind of body than with a new way of learning. The industry is moving from a world where every motion had to be scripted in advance to one where behavior can first be learned, then transferred to another robot, and then improved in deployment. This isn't yet the household android science fiction promised, but it is a clear engineering trajectory, and it's why robots have once again become a big bet for researchers and investors.