A Tough Robot Test
A startup has developed a universal AI system for controlling robots. It can adapt to changing conditions. For example, it helps a robot continue to move with damaged limbs.
The startup Skild AI has solved a fundamental problem in robotics in an original way. They trained a single AI brain to control 100,000 different robots in a simulation. Then they started sawing off their legs (with a chainsaw!), locking their wheels, and putting them on stilts. The robot, even with its lower leg cut off, learned to walk in 7 seconds.
The logic is ironclad. A standard AI for a robot is like a student memorizing the answers to an exam. It works perfectly in familiar conditions, but change one parameter and everything falls apart. Skild AI's omnibrain can't remember the solution for a single body; it has to find a strategy for all 100,000 at once.
The tests are sadistic. They lock the robot's knees, and it learns to walk on three legs in 2-3 seconds. If the wheels lock, it switches from driving to walking. If you put it on stilts, raising its center of mass, it adapts its gait in just a few steps. A specialized controller, under the same conditions, simply falls over and flips over.
After a millennium of simulated time, we've created a brain that adapts to changes in milliseconds or minutes. Skild AI claims these are the first sparks of intelligence in the physical world. Robots that will survive any damage.
