One of the most important robotics ever done was when some scientists released a robot into a large shopping mall to see how children would react. Beyond the feeling of curiosity about what they were seeing, the children tried to kick and hit the robot.
The conclusion of this experiment was that perhaps they should let robots go free to see what movements they would make and how they would navigate the world. Researchers at the University of Southern California have concluded that letting robots hit objects with their hands will help them learn to grasp those objects better.
The experiment was conducted entirely in simulation. The environment simulates variables such as friction, and a robotic arm tries to grasp an object using different grips. If it doesn't stumble, the system counts it as a win, and if it makes a mistake, the system counts it as a loss. Over many trials, the robot learns what constitutes a strong grasp.
But this is where the so-called “human actor” comes in, a kind of additional signal. If the robot manages to grasp an object well, the human uses an “interface” to click on the object it is holding and apply a force in a specific direction.
However, there are a few things you need to know. A simulation is a necessarily imperfect model of the real world. There is no way to make exact copies. So transferring what a robot learns in simulation to a physical robotic arm is still very difficult.
