How OpenAI, Google, and Figure Are Changing Robot Training: A Brief History of the Approach

Robotics is undergoing not just another cycle of hardware development, but a fundamental shift in the very logic of learning: machines are increasingly trained on data, simulations, and large multimodal models rather than being programmed by hand. For a long time, the industry had a notable gap between ambitions and results. Engineers dreamed of science-fiction-level robots — universal assistants capable of moving through the ordinary world, understanding people, and safely performing dozens of tasks.

In practice, the market has lived for decades largely on the back of narrower systems: robotic arms for factories, warehouse manipulators, and consumer devices like robot vacuums. The reason was not a lack of mechanics, but the fact that robots struggled with unpredictable environments. Any deviation from the script — a new object, a different viewing angle, changed lighting — quickly broke their behavior.

The first major shift came with learning through trial and error. In the 2010s, roboticists began actively using reinforcement learning and simulators, where machines could train millions of times without risking breaking expensive equipment. One notable example is OpenAI's Dactyl robotic hand, which was trained in simulation to manipulate a Rubik's cube, then transferred the skill to the real world. The key idea was domain randomization: during training, the parameters of the environment changed — friction, object mass, camera images, and other details. This way, the robot learned not one perfect scenario, but robust behavior across a wide range of conditions.

Parallel to this, learning from demonstrations developed: robots began to be taught by showing them human actions rather than manually scripting every rule. But even that was not enough. Simulations accelerated learning, but the gap between virtual and physical environments didn't disappear, and real data remained expensive and scarce.

So the next stage began when robotics adopted approaches from the large models boom. Instead of separate systems for individual tasks, researchers started assembling large datasets of trajectories, images, text instructions, and actions. At Google, the RT-1 model was trained on 130,000 episodes collected by a fleet of 13 robots covering over 700 tasks.

And RT-2 went further: it combined robotic data with web data from vision-language models, so the robot could not only repeat familiar movements but also make broader inferences from natural language commands.

This same turn is visible in projects attempting to scale learning across different types of machines. In the Open X-Embodiment dataset, researchers collected data from 22 robots from 21 institutions and described hundreds of skills so that models could transfer experience between platforms rather than starting from scratch for each new manipulator. On top of this rests the idea of foundation models for the physical world: a single base intelligence that is fine-tuned for a specific body, gripper, or work environment. Startups like Figure, Apptronik, Covariant, and Physical Intelligence are building not just research but also commercial strategies around this.

Against this backdrop, the market has revived. In 2025, companies and investors invested 6.1 billion dollars in humanoid robots — four times more than the year before.

The money flows not because robots can already do everything, but because a more plausible path to their training in real environments has emerged. Yet universal robots remain far off. Machines still struggle with long sequences of actions, fine manipulation of multiple objects simultaneously, work in cluttered spaces, and safe interaction with humans without pre-prepared scenes.

Unlike language models, a robot cannot simply make an error in text: its mistake means a falling object, a breakdown, downtime, or risk to humans. So the industry moves forward via a hybrid path: more simulations, more real data from fleets, more general models, but also more constraints from safety and economics.

The main conclusion is simple: the breakthrough in robotics right now is connected not so much to a new type of body as to a new learning model. The industry is transitioning from a world where every movement had to be prescribed in advance to one where behavior can first be learned, then transferred to another robot, and further improved in operation. This is not yet the promised android from fiction, but already a clear engineering trajectory that has made robots a big bet again for researchers and investors.