The South Korean company revealed new details about how the Atlas robot learned to lift and carry heavy industrial loads using artificial intelligence, reinforcement learning and millions of digital simulations and perhaps the most impressive thing is not just the weight, it is the way the robot learns to handle it. In the new tests, Atlas appears carrying industrial bars weighing more than 45 kg like refrigerators, while continuously adjusting the balance of his own body in real time.

The robot rotates its torso, crouches, calculates force distribution and reacts to internal weight changes while walking, all without relying primarily on camera vision. According to the company, Atlas uses something called proprioception, a kind of internal body awareness, it is the same principle that humans naturally use to know where their arms, legs and center of gravity are without having to look directly at them.

The robot feels its own body as it moves, to teach this the engineers did not manually program each movement, instead they used reinforcement learning. Atlas spent millions of hours training in virtual simulations run on GPUs, repeating movements countless times under different conditions.

The researchers altered the weight of the objects, the level of ground friction, the force needed in the hands and even small external perturbations to force the system to continually adapt. Every time the robots correctly completed a task while maintaining balance and control, it received an algorithmic reward, little by little it began to discover the most efficient movement patterns on its own.

The above represents an enormous change in robotics, because for decades industrial thefts depended on extremely predictable environments, everything had to be fixed, controlled and repetitive, small physical changes were already enough to generate failures. We are now slowly beginning to build machines capable of physically improvising in the real world. Boston Dynamics itself affirms that one of the greatest evolutions of the new Atlas is precisely to reduce the famous gap between simulation and reality, the historical problem where robberies work perfectly in the virtual environment, but fail when faced with the unpredictable chaos of the physical world.

And there is something interesting in all this, the athletic movements that previously seemed like just marketing, somersaults, extreme balance, are now beginning to make sense industrially, according to the company, these skills help develop slip recovery, dynamic stability, thermal resistance and body adaptation for real work environments, that is, perhaps those impressive videos were never just spectacle, perhaps they were disguised training to prepare machines capable of replacing human work on an industrial scale.