The Robot Exercise

Overview

In this exercise, four students simulate a robot. One acts as the robot's vision system, one as its brain, and two as its arms. The task of the robot is to stack one box on top of another. The boxes should be sufficiently large or heavy that they cannot be picked up with one hand.

When?

We have used this exercise at two different points in an undergraduate AI course: the first day of class or at the point in the course when we switch from "brain-in-a-box" topics like search and planning to topics dealing with sensing and acting, like vision, NLP, and robotics. It works well early in the course, as both an ice-breaker, and a demonstration that what seems easy is not necessarily so. It could also be used in a general introductory CS course as an introduction to AI.

Before

Before beginning, explain to the students that they will be considering how an intelligent robot interacts with the world. Describe in general terms current capabilities of vision and robotic systems. You may want to use illustrative anecdotes from the recent literature or from the AAAI robot competitions.

Describe the task to be accomplished: given two boxes on a table, we wish to stack one on top of the other. Ask for volunteers and choose/draft four students. It is best to choose the brightest and most outgoing one to be the brain as this task is tough and frustrating even for a really good student. The arms are blindfolded, and the brain is seated facing away from the table on which the boxes are placed. Here are the ground rules:

• The brain can talk to any or all of the other three robot parts.
• The arms follow the directions of the brain. They should pretend not to hear the vision system.
• Explain that realistically the arms should be directed in terms of joint angles and forces, but that we will pretend that they are exceptionally smart arms and can understand directions in terms of moving their hands a give distance in a given direction, up, down, left, right, forward, and backward.
• The arms can answer yes/no questions from the brain and they can tell whether they are touching something or not. They cannot however distiguish the table from the boxes or the other arm.
• The vision system cannot volunteer information, but it can answer (as literally as possible) any questions posed to it by the brain. Point out that this vision system is also very smart in that it can easily distinguish the table from the boxes and can estimate distances fairly well (depending on the student).
• The vision system is allowed to move around to get a better view of the scene.

The Exercise

It's best to allow most or all of a class period for this exercise, as it's likely to take 30-60 minutes to get the boxes stacked. While the students are working, you may want to step in occasionally to ask the brain to explain what it is doing or comment on what is happening. If the students get frustrated or stuck replace them with new volunteers (be sure to change the box configurations in this case).

After

• Coordinating the arms is hard.
• Visual feedback, force feedback, etc. are important, and a big reason why this problem is difficult is that the feedback loop is indirect and inaccurate.
• World knowledge is important.
• Problems arise from different coordinate systems, for example box-centered vs. arm-centered.
• The brain often starts using his/her hands to demonstrate how the arms should move or to clarify a question. This brings up a chance to talk about model-based communications and reasoning.
• The difficulty of constructing a plan without knowing the preconditions is quickly seen. Soon, the brain figures out that it must monitor the plan execution consistently, and replan as necessary. Something unexpected always goes wrong (e.g. both hands are touching the box but not on opposing sides or not in the same horizontal plane; or the top box is above the bottom box but its center of mass is not), which points out how hard it is to forsee all contingencies.

Follow-up Assignment

Back: Using Robots in an AI Course