DeepMind teaches mini-robots how to play football
The Google subsidiary DeepMind uses artificial intelligence to train robots for a one-on-one football match. The training also involves the researchers pushing the funny little bipeds.
The company DeepMind, which specialises in artificial intelligence, is teaching small robots, each of which has 20 joints, how to play football. A team of researchers is investigating in more detail whether it is possible to develop sophisticated and safe movement skills in a cost-effective way using deep reinforcement learning. And whether these skills can be combined to form complex behavioural strategies in dynamic environments.
The robots were trained for a simplified one-on-one football match. In the first videos, the researchers show the current state of play. The results are entertaining, sometimes seemingly nasty, scenes - such as this push training session:
First, the team practised individual skills in isolation in a simulation before the robot itself was allowed onto the real pitch, initially on its own. The training included things like recovering quickly from falls, walking, turning, kicking and more. Stable and efficient transitions between individual actions were also trained.
The robots in the following video are already trained and compete one-on-one. In addition to the movements, they have also learned complex behaviours. They have a basic understanding of strategy and are also able to anticipate ball movements. This enables them to block opponents' shots. The behavioural training also took place in a simulation before the software was transferred to the robots.
The AI training, together with minor hardware modifications, is said to have ensured that the robots move more safely and effectively compared to basic programming, without sacrificing dynamics and agility. Thanks to the AI training, they are said to take 63 per cent less time to stand up, run 156 per cent faster and kick the ball 24 per cent quicker.
DeepMind is satisfied with the preliminary results. Their analysis shows that using deep reinforcement learning for robot training is a promising approach. The team wants to continue and, in particular, investigate what the robots can learn from their camera image and whether it is possible to better predict various aspects of the game.
Cover image: Screenshot YouTube video / OP3 Soccer Team, DeepMind
I find my muse in everything. When I don’t, I draw inspiration from daydreaming. After all, if you dream, you don’t sleep through life.