Grounded language can be demonstrated with a vocabulary game. The game engien prsents a camera picture and a human has to enter the correct word. In case of a warehouse robot game, such a quiz consists of a BNF grammar which holds only the possible words, and a picture which contains all the possible objects and actions.
BNF Grammar
<word> ::= <objects> | <actions>
<objects> ::= "wall" | "door" | "charging station" | "human obstacle" | "box" | "shelf" | "robot obstacle"
<actions> ::= "move straight" | "turn left" | "turn right" | "u-turn" | "stop" | "speed up" | "speed down"
Picture
The interaction with the game is easy to explain. The random generator selects the picture with the "door" and the human has to enter the word "door" into the input field. The game engine recognizes that the answer is correct and rewards the human with a +1 score.
The harder to explain question is why such a vocabulary game is relevant in robotics. It has to do with the previously mentioned BNF grammar. The grammar provides a minimal state space for the game. Its not important what sort of camera is used by the robot and which programming language is available e.g. C++, but only the BNF grammar counts because it holds all the possible game assets. In the concrete case the BNF grammar is stored in a text file with 251 bytes and contains 14 different words. These words are providing the conceptual space for the robot.
Suppose a human operator submits a command sequence like:
1. Enter the door, bring the box the shelf
2. then move to charging station
3. if there is a human opstacle or a robot obstacle, stop!
Such a command sequence is valid for the robot because it is based on words from the BNF grammar. The robot converts the command sequence into actions. This allows a high level human to machine interaction
No comments:
Post a Comment