Train robots to perform simple tasks in a wide range of environments using sim-to-real reinforcement learning

A team of roboticists at the University of California, Berkeley reports that it is possible to train robots to perform relatively simple tasks by using simulation-to-realistic reinforcement learning. doing.In their research published in the journal science roboticsa research group trained a robot to walk in an unfamiliar environment while carrying various loads without falling.

Over the past few years, roboticists have used a variety of techniques to train robots to move efficiently and quickly in a variety of environments. But as the researchers behind this new effort point out, there aren't many useful applications for such robots. They suggest that robots that can perform routine tasks slowly but efficiently would be far more useful. To achieve this goal, they turned to simulation-to-realistic reinforcement learning.

The technique involves exposing a robot to billions of samples in a simulated environment to train a simulated version of the robot to perform the desired task. This method also involves using a reward/penalty system as part of the robot's training. If the robot does something correctly to achieve a goal, it is rewarded by receiving a “1”, for example. However, if you do something wrong, it will return “-1”. Over time, your performance improves as you try to increase the number of rewards.

The research team used this approach to navigate paths along sidewalks in unknown parts of town, recover after being repeatedly attacked with large balls, overcome physical restraints, and cross dangerous materials. We trained a robot called Digit to walk. They trip, carry backpacks, carry bags of trash to the dumpster, and use totes to carry personal items.

Researchers suggest that simulation-to-realistic reinforcement learning could be used to train robots in real-world environments such as homes, offices, and factory floors. The idea, they say, is to make robots more useful.