Have you seen AI wrestling in sumo style? Can AI agent kick a goal? Interesting?
Yes, it was as interesting as it can be. Open AI, a non-profit organization tried to train Ai agents sumo wrestle and Kicking ball into goal.
These AI consist of two agents and both were seeking to maximize their reward. Initially, each agent was rewarded for moving around its environment, exploring its surroundings. Researchers then started narrowing the reward parameters and made them specific to the goal.
In case of Sumo Wrestle, both agents were rewarded for moving or exploring the sumo ring, then rewards made zero in orders to give rewards only for winning and losing. Despite the simple rewards, the agents learn subtle behaviors like tackling, ducking, faking, kicking and catching, and diving for the ball and the specific kick the opponent out of ring.
Each agent’s neural network policy is independently trained. In one case agent was trained individually with self play while agent need to stand against unknown “wind” force that disturb it. The agent managed to stay there despite it never seen that “wind” force.
After each iterations the wrestling skills got better and better. They even taught themself to fool the opponent and managed to fool and finally kick it out of the ring. The same approach worked for other challenges like soccer and tackling. While these are cool tricks, it’s important to remember that all of these behaviors simply reflect optimized solutions to myriad calculations. Sure, they look like humanoids, but it’s all math.
AI is moving very fast towards its height as much of the research is being done in AI. the results are simply awesome.
