Send an army of robots.
According to an initial report from Cable.
The new tricks learned in the simulation could soon see performance in a neighborhood near you.
The robots in the simulation took control of the step and blocked the navigation
The simulated army was developed by Swiss-based researchers from ETH Zurich, in addition to the engineers at the chip company Nvidia. Together they used meandering bots a simulation called ANYMals to overcome difficult obstacles for robots, such as steps, slopes and sharp drops carved into a virtual landscape. Each time a robot solves a navigation problem, researchers make it more difficult by putting the algorithm into an insanely unforgivable puzzle whose sole purpose is to teach its digital guest how to overcome the insurmountable, achieving a level of sophistication unseen in AI mobility.
Graphically represented, the ensuing drama unfolds like an army of confused ants writhing through a giant sea of geometric madness. As they go through training, robots master walking up and down without too much struggle. But the slopes threw them a curve. Few could grasp the basics of gliding. But after the final algorithm was moved to a real version of ANYmal, the four-legged canine robot with sensors equipped in its head and the robot’s movable arm moved successfully on blocks and stairs, but had problems working at higher speeds.
An army of robots with negative AI feedback
Researchers do not blame the algorithm. Instead, they believe that the lack of correspondence between the way sensors perceive the real world and the virtual world creates problems with coordination. But this kind of accelerated robot training can accelerate the learning curve for robots and other machines to learn a wealth of skills, from sewing clothes and harvesting to sort packages in a colossal facility on Amazon. The project also confirms the importance of using simulation to improve artificial intelligence (AI) capabilities. “High-level, very fast simulation is a really great thing,” said Professor Peter Abil of UC Berkeley in Cable report. Abbeel is also a co-founder of Covariant, a company that uses AI in simulations to train robotic weapons in the art of sorting objects for logistics companies.
According to Abbeel, the work of researchers from Switzerland and Nvidia with robotic algorithms “has achieved some good accelerations”, according to the report. AI has come a long way and now it can improve the ability of robots to perform tasks in our daily lives that are not easily translated into software. Capacity to grip awkward, strange and slippery surfaces, for example, is not something you can reduce to a few lines of simple code. That’s why 4,000 simulated robots trained with reinforcement training, which is an AI method that takes its cue from the way animals learn by positive and negative feedback. As the robots move their legs, the evaluation algorithm monitors how this contributes to the robot’s ability to continue walking and adjusts the control algorithms to adapt as the movement continues. Nvidia’s specialized AI chips supported the simulations, allowing researchers to train an army of robots for a hundredth of the time they would otherwise require. We finally got to the beginning of self-learning robots and from combining supportive training with the latest advances in AI, the limits of robotic motion can approach the limits of the physical world.