The evolution of robotic learning.

With less and less time invested

Imagine showing a machine completely blank, without any instructions on how to move the legs and arms and seeing in front of your eyes how it discovers on its own how to balance its own body, first come the falls, then the clumsy movements, then small hesitant steps and finally a stable, almost natural walk. What looks like a birth scene is actually one of the biggest revolutions in modern robotics.

The era when engineers programmed each joint manually is coming to an end. At this point, artificial intelligence learned to dominate its own body and why this represents a historic change for machines. Before, teaching a robot to walk meant calculating thousands of mathematical variables, every angle, every tilt and every balance adjustment had to be programmed by hand, but that all changed with reinforcement learning, instead of receiving detailed instructions, the machine learns through trial and error.

Experience random movements, fall countless times, and receive virtual rewards when you manage to move forward or maintain your balance. Little by little, artificial intelligence discovers on its own which strategies work best. To accelerate that learning, companies like Figure, AI, and Boston Dynamics created a parallel world inside supercomputers. Millions of digital copies of robots live in accelerated simulations, tripping, slipping and facing obstacles for thousands of virtual years and only in real hours.

When AI is placed on physical hardware and finally turned on, it already comes into the world with enormous accumulated experience. This method is known as Sim to Real, but researchers at the University of California at Berkeley and Google Brain decided to follow an even bolder path. Instead of training inside computers, they placed the robberies with it directly in the real world, walking on carpets, mattresses and uneven surfaces, the machines learned to get around in a matter of hours.

Robots began to discover for themselves how to interact with the physical world.

The secret lies in proprioception, a kind of internal sense that allows robots to perceive weight, pressure and balance, in the same way that the human brain receives information from muscles and joints. The speed of this evolution is impressive, in 2020, Google's Rainbow Dash quadruped took only 3 and a half hours to learn to move on its own, in 2021, the bipedal robot Cassie could already walk, crouch and carry loads using only neural networks.

The following year, researchers broke another record by teaching a quadruped to walk in the real world in just 20 minutes, but it was in 2026 when the machines definitively left the laboratories; the humanoid Agibot completed an autonomous journey of more than 106 km between Suzhou and Shanghai, walking for about 56 hours. Meanwhile, Boston Dynamics' new Atlas went through weeks of training using reinforcement learning and began executing extremely complex tasks, from inspired movements and sports to carrying loads exceeding 45 kg, constantly adapting its own balance.

Perhaps the most impressive achievement of this new generation of machines is not strength or speed, it is the ability to learn. For the first time in history, robots stopped just obeying programmed movements and began to discover for themselves how to interact with the physical world. And if artificial intelligence has learned to control its own body, the inevitable question is, how long until it surpasses human agility?