Game-trained robots

We have computer games: these create a simulated universe with which the player interacts; the game evaluates each player's interaction based on goals coded into the game. We've tried various ways to construct control systems for robots; in the case of some kinds of machine-tool, having a proficient operator of a non-robot tool operate the tool - while monitored - has been used as a way to provide the robot with its instructions for how to do its job. It seems to me that there is scope for a marriage of these approaches. A game could record what its players did, and how well they fared in its simulated universe, and use this as the raw material for controlling a robot.

In particular, in the field of remote presence - whether in space, the deep ocean or other dangerous domains - there's scope for quite interesting games. The scenario users play in can be embelished - most obviously, consisting almost entirely of the interesting parts, i.e. the ones where the robot needs some help, but also including artifice in so far as it shares, with the robot's reality, the essential elements of the problems it's not easy to pre-program the robot to solve. That can be used to make the game interesting enough that plenty of people will want to play. It can also be used as the basis for separating (essentially in the same pattern as the levels in many games) the more competent players out from the general mass.

Honestly robust guarantees of privacy can be offered to players while still collecting information potentially useful to the robot (but one may need some form of waiver of any right to be paid for use of those ideas). As long as large numbers of folk play, plenty of information can be collected on how well various strategies work. The results of this can be compared to the real problems the robot needs to solve; in so far as the results match reality, players can gain credit (in the game) for their strategies proving helpful; in so far as what the game thought was good turns out not to work well in practice, the game can be refined.

Real pieces of equipment can then be driven by software that has learned from the players of a game. When the equipment does something valuable (or merely newsworthy), the game can tell players who had (not necessarily uniquely) contributed ways to cope with reality whose application contributed to that; whether this contributes to their score - whether the primary game score or some secondary kudos score - is incidental, since they can feel a sense of involvement through it. Less valuable similar pieces of equipment (in, for example, safe or controlled environments) can be used to validate the game's simulation of lower levels.

Such real pieces of equipment can also be used as sources of data by the game's simulation of reality - the player can never be entirely sure how far the simulated scenarion being played is, in fact, a live feed from the actual piece of remote equipment. One way to deliver their kudos score is by showing them the actual real-world robot's record of what it did and how well it did - maybe annotated with the simulator's opinions of where the player's deviation from the robot's behaviour might have been an improvement. The real robot's performance can also be used as a calibration reality for the scoring system in the game.

The optimal case for such game-trained robots is space, where the robot can be in permanent communication with the internet (hence with the players) despite being at the wrong end of a latency that means a direct human controller is not a viable option. Similar approaches can be used for deep-sea - and other dangerous environment - robots even though they are amenable to real-time control by a real live human. They can be prepared with many response strategies derived from game players, which they might chose to use in the short term in preference to inaction when the real live human controller hasn't done anything about a situation the robot has already determined to be a problem. Equally, a robot could be controlled by some form of automatic voting (most of the players to whom the game is offering the robot's present actual reality chose to run away - OK, let's run away) among many game players, rather than by a real live controller; the players who responded differently get to see a simulated game reality past the point where they diverged from what the majority chose. Of course majority might be weighted in some manner that favours the choices of those players who've tended, in simulated realities, to do well.

See Also

As XKCD 1356 points out, the Kerbal Space Program game has taught its users far more about orbital mechanics than any school ever did. It would be interesting to see whether any of those playing the game have come up with cunning orbital manoeuvres that the scientists at the world's various space agencies hadn't thought of. Increasing, by several orders of magnitude, the number of folk eager to solve a problem has a good chance of delivering better solutions, even if the original few were way smarter than the many new-comers.

Written by Eddy.
$Id: rome.html,v 1.2 2004/08/15 18:03:12 eddy Exp $