from Hacker News

Imagine a GPT-X model that can understand concepts like algorithms and how to compile lists of steps based on a strict set of available instructions. This would ultimately fill the gaps of writing arbitrarily complex code to allow a robot to understand instructions in human language.

Human: Hey Robo1000, go to the kitchen, and get me a bag of potato chips out of the pantry.

Computer instructions inside Robo1000: Given that I am a robot, with the following instruction set:

   1) determine what direction I'm facing

   2) determine if there is an obstacle in my way

   3) go right

   4) go left

   5) go forward

   6) go backward

   7) open door

   8) pick up object

   9) identify object

   10) determine what room I am in

   11) determine where my destination is relative to my current location

   12) identify the location of people in a room

   13) identify which person made the request

   14) etc, etc...

  Assumptions:

    1) If there is something in my way I must find a way to navigate around it

    2) etc. etc..

  Given that my instructions are to: "go to the kitchen, and get {name of as yet unidentified person} a bag of potato chips out of the pantry", generate the shortest algorithm that I can follow repeatedly that will make it possible for me to accomplish this task.

  Generated list of instructions to accomplish task:

  1) Determine who made the request

  2) Determine what direction I am facing

  3) Determine where my destination is relative to my current location

  4) Turn left...

  5) ... etc...