In most of the scenes we have seen lately with robots, where they perform spectacular movements, somersaults, acrobatics or martial arts, there was a small detail hidden behind the demonstration, many times that was just movement programmed specifically for that presentation, that is, the robot had not really turned on the skill, although very impressive, what it did was reproduce a previously trained sequence, but now Chinese researchers may have taken an important step to change that.

The Beijing Institute of General Artificial Intelligence, in collaboration with Unitree Robotics, presented a new system called Omni Extreme, the proposal is to allow a single artificial intelligence to control a large library of complex movements. without losing performance. In humanoid robotics there is a classic problem called the generality barrier. This is when researchers try to teach many different movements to a robot, jumps, turns and acrobatics, the system ends up generating safer and generic movements that usually fail in really dynamic situations.

The Omni Extreme attempts to solve this with a two-stage approach, training and simulation, using a high-capacity model that learns by observing several specialists in motion. Instead of creating one skill at a time, the system learns to imitate a wide variety of behaviors simultaneously. Refinement with reinforcement learning that adjusts the movements considering the real physical imitations of the robot such as the torque of the motors, the speed of the actuators and the energy consumption.

The goal is to ensure that what works in the simulator also works in the real world, the results highlighted in more than 150 tests carried out on Unitree Robots, the system managed to execute different categories of movements with high success rates, jumps and turns, more than 96% success, martial arts movements, above 93%. and breakdancing in its robotic rate close to 86%.

If systems like Omni extreme continue to evolve, robots of the future will be able to learn hundreds or even thousands of different physical skills, well, up to this point we saw that robots really learn.