Please carefully read the instructions below.
In this experiment, you will be training the Fetch robot (the big robot in the video on the right) to serve drinks.
Setting. Fetch is operating as a chef in a restaurant. It has a customer (the small robot) who ordered a burger combo. Fetch has already served the burger and potato chips. It is also cooking some fruit sauce on the stove for the dessert, but meanwhile it needs to serve a drink before the dessert. It has a few different ways of bringing the drink.
Goal. Your goal is to train Fetch to bring the drink in the manner that you prefer. Below is a list of factors that the robot takes into account when bringing the drink, which you can optimize based on your preferences. Please read them carefully while also considering what you would prefer.
Factor 1. Which drink should Fetch serve for the burger combo? From left to right, the drinks are: orange juice, spring water, and milk.
Factor 2. In what orientation should the pan be? Having the handle close to Fetch makes cooking easier, but you may also prefer having a clear path behind the stove in case you prefer the drinks to be brought from behind the stove as opposed to moving over a hot pan.
Factor 3. Should Fetch move the drink behind or over the stove?
Factor 4. If Fetch decides to move the drink over the stove, how high above the pan should it move?
Factor 5. Overall, how fast should Fetch move? A fast service is often desirable by the customers, but slow robots may feel safer.
Procedure. Based on these five factors, your task is to train Fetch with respect to your own preferences. For this, you will be responding to various forms of questions. Let's now go over them one by one:
Scale Questions: Most of the time, you will be given two trajectories to compare. You will then use a slider bar (see the figure below) to answer "how much you like one trajectory relative to the other?". Moving the slider bar to the right means you like the trajectory on the right more. Similarly, moving it to the left means you like the left trajectory more. You can also leave it in the center if you are indifferent between the trajectories. Please NOTE that slider bar is NOT a measure of uncertainty. Rather, use it to reflect how much better one trajectory is than the other. As an example, suppose you are given two trajectories that are equivalent in all aspects except one is slightly faster than the other. In such a scenario, you could be very confident that one trajectory is better than the other. However, you should move the slider bar only slightly to indicate the difference between the trajectories is small.
- Choice Questions: You will also be asked a sequence of choice questions. In these questions, you are again shown two trajectories and given a slider bar. However, as opposed to scale questions, you cannot give intermediate responses to the choice questions. You will either choose a trajectory (by moving the slider bar to the end right or left), or you will leave it in the center to indicate you are indifferent between the trajectories.
- Ordinal Questions: Occasionally, the interface will show you a single trajectory and ask to confirm it fits your preferences. You will respond to these questions with one of the five options: 1- Strongly Disagree (it does not fit), 2- Disagree, 3- Neutral, 4- Agree, 5- Strongly Agree (it fits).
Interface. For scale and choice questions, there will be a "Restart Videos" button at the top that can be used to restart the two videos to sync them at the same time for an easier comparison between the two behaviors. Similarly, there will be a "Restart Video" button for the ordinal questions which will restart the trajectory.
Now, please complete the 2-question quiz below to start the experiment.
- What does it mean if you move the slider bar with a large margin to the right in a scale question?
- Which of the following is true?