Artificial intelligence has physically beaten humans as well as out-thinking them in what is claimed to be a new breakthrough.
Researchers in Switzerland created an AI robot that learned in just six hours to master a popular game of physical skill.
They claimed that its success through “deep reinforcement learning” was pushing back the boundaries of artificial intelligence, out-performing humans physically rather than just cerebrally.
An AI robot named CyberRunner was tasked with learning how to play the popular labyrinth marble game.
It is a game of physical dexterity whose goal is to steer a marble from a set starting point to the end point, while stopping the ball from falling into any of the holes on the labyrinth board.
The robot made initial mistakes but then learned to complete it in 15 seconds.
Humans use two knobs that change the orientation of the board to direct the marble around it, using fine motor skills, spatial reasoning abilities and experience.
But it can take them considerable time to be able to successfully complete it.
The researchers stressed that CyberRunner applied advances in model-based reinforcement learning to the physical world and exploited its ability to make “informed decisions about potentially successful behaviours by planning real-world decisions and actions into the future”.
It learned through experience, capturing observations while playing the game and receiving rewards based on its performance.
These functions were done through the “eyes” of a camera looking down at the labyrinth.
A memory was kept of the collected experience.
Using this memory, the model-based reinforcement learning algorithm learned how the system behaves, and based on its understanding of the game it recognised which strategies and behaviours were more promising, according to the researchers.
So the way the robot used two motors, its “hands,” to control the board was continuously improved as the algorithm was running concurrently to the playing of the game.
The learning process on the real-world labyrinth was said to have been conducted in 6.06 hours, with 1.2 million time steps at a control rate of 55 samples per second.
The AI reportedly robot outperformed the previously fastest recorded time by a human, by more than six per cent.
Trumpeting the CyberRunner, Professor Raffaello D’Andrea, at ETH Zurich public research university, said: “We believe that this is the ideal testbed for research in real-world machine learning and AI.”
He claimed further: “Prior to CyberRunner, only organisations with large budgets and custom-made experimental infrastructure could perform research in this area.
“Now, for less than 200 dollars, anyone can engage in cutting-edge AI research. Furthermore, once thousands of CyberRunners are out in the real-world, it will be possible to engage in large-scale experiments, where learning happens in parallel, on a global scale. The ultimate in Citizen Science!”
Perhaps most strikingly, CyberRunner discovered shortcuts as it learned, according to the researchers.
It found ways to ’cheat’ by skipping certain parts of the maze.
The lead researchers, Thomas Bi and Prof D’Andrea, intervened to explicitly instruct it not to take any of those shortcuts.
The research on the www.CyberRunner.ai was carried out at the Institute for Dynamic Systems and Control at ETH Zurich.
Rishi Sunak last month held an artificial intelligence safety summit in Bletchley Park, the once top-secret home of the World War Two Codebreakers.
Artificial intelligence chiefs and experts including Elon Musk and some world leaders attended the gathering which focused on “frontier AI” which offers the greatest potential to deliver economic, health and other benefits but also could pose an existential risk to mankind.
