Sunday, November 10, 2013

We'll Need Virtual Reality for Robots

Within several decades we presumably will have robots capable of walking around among us, acting as helpers. One important issue is how the robots will be trained. Let us consider the case of a robotic policeman, rather like the hero of the movie Robocop, but entirely electronic (rather than the man/machine blending depicted in that movie). How could such a robotic policeman be trained?

One approach would simply be: introduce the robot into a working situation, note any shortcomings, and correct as necessary. But this would be extremely dangerous. Let's imagine an example of how things could go very wrong. A robot policeman encounters a group of young boys playfully punching at each other, without actually trying to hurt each other. This is something that young boys often do. But the robot policeman interprets the arm-swinging activity as one person assaulting another. The robot policeman then tells the boys to freeze. One boy takes out a black colored cell-phone, but the robot misidentifies the object as a gun. Thinking it is in danger, the robot then pulls his gun and shoots the young boy.

Or let's imagine a family that buys a household helper robot. The father arrives at home in his car with a big heavy box containing a new television. He summons the robot, and asks him to carry the big heavy box from the car, and put it in the living room. The robot picks up the box, and takes it into the living room, depositing it on the living room floor. Unfortunately, just before he places it down on the floor, a two-year old toddler crawls below the spot the robot selected to place the big heavy box. The young child is tragically crushed by the huge heavy box.

Clearly it won't be good to allow these types of tragic incidents to occur. What will we need is some kind of way to train android robots before they are let loose in the real world. How could such training occur?

One possibility is to create a huge “sandbox” training center that would mimic conditions in the real world. The training center would be populated by two types of android robots: the robots that were meant to be introduced into the human community, and other robots intended only for the training center, robots designed to simulate human beings. Software engineers use the term “sandbox” to mean an environment in which a software application can be tested, without causing harm in the real world. A training center designed to simulate a real-world environment would be kind of a giant “sandbox” for robots to be tested in. If a robot made a mistake, he might damage or destroy one of the training robots designed to simulate humans. But no real human would be hurt.

Such a sandbox training center would probably be a good idea, but it would be very expensive to create such centers. It would be particularly expensive to create training robots designed to simulate human beings.

Is there some less expensive way that robots could be trained? There is: virtual reality. It seems that virtual reality training might be a vital component of the training of advanced robots.

 A robot using virtual reality for training

Here is how it might work. A virtual environment could be created to train robots, an environment similar to the virtual worlds created for video games. A robot could then be allowed to train in such an environment. Instead of interpreting visual data from the real world, the robot would interpret visual data supplied by the virtual reality system. If the robot made a mistake, it would never cause real-world death or destruction, but merely death or destruction in the virtual world.

We can imagine, therefore, a sequence that could be used to train new android robots:

  1. First, a new model of android robot would spend extensive time training in a computer generated virtual reality environment. As it was training, any problems in the robot's responses would be noted, and corresponding corrections would be made in the robot's software.
  2. If it passed these tests, the new model of android robot would spend time in a physical “sandbox” environment consisting only of itself and other robots designed to simulate humans. As it was training, any problems in the robot's responses would be noted, and corresponding corrections would be made in the robot's software.
  3. Finally, if it succeeded well enough in the physical “sandbox” environment, the new model of robot would be sent out into the real world to interact with actual humans.

The same approach may become an option in the future for training human beings for certain types of jobs that must be performed properly, or else people will die. For example, imagine someone training to be a doctor.

  1. First, the student could spend extensive time using virtual reality training programs, interacting with computer-generated patients with computer-generated symptoms.
  2. Then, the student could spend time in a physical “sandbox” environment containing only robots who simulated sick human beings. We can imagine robots deliberately manufactured with fake cancers and fake bone fractures, which would make great tools for training surgeons.
  3. Finally the student could be let loose to begin treating actual human beings.

No comments:

Post a Comment