Realistic gaze behavior of CCHD characters
 

Hi all,

I'm planning an experiment in which I want to establish eye-contact between a participant and a CCHD character. I can make the character gaze at the participant all the time, but that does not seem very realistic. Has someone ever done a similar experiment with more realistic gaze behavior (e.g. glancing away)? If so, I would be glad to learn more about it!

Hey Frank,

I have done a lot of work with eye gaze. Do you use any eyetracking in your experiment? If yes, you can reuse participant's gaze data and modify it a little bit to animate your avatar. For example, gaze following is the most natural thing in the world, and you could make your avatar follow the participant's gaze 70% of the time, and break eye-contact for 30%, where avatar is looking at something that interests it. Otherwise, I suggest prerecording somebody's natural eyegaze and using this to animate the eyes, since except from some robotic work, nobody has really managed to program a perfect simulation of natural gazing.

And don't forget blinks! They happen every 5-10 seconds, and should last between 100 and 500 ms.

Hi Iva,

Thanks for your reply! I do not use eye-tracking in my experiment (and don't have the equipment to do so), although I want participants to experience eye contact with a virtual agent. The way I'm accomplishing this right now is to have the agent face the participant, and the agent either looks straight ahead (thus making eye-contact when the participant looks at the agent), or averts its eyes. Have a look at the script attached to this message to see how I handle blinking and looking away behavior. For the looking away behavior, I take a random float between -15 & +15 for the yaw, and -10 & +10 for the pitch. The eye moves every 2-6 seconds. I don't know if these settings produce humanlike looking away behavior, but I'm pretty pleased with the result so far. If you have any suggestions to improve the parameters, let me know.

Do you have any reference for the percentages of maintaining and breaking eye-contact? If have seen some, but they distinguish between listening to someone and talking.

The blinking behavior I use is from here. Do you have a reference for the time between blinks and their duration? With this information, I could improve the eye-blinking behavior of my agent.

Thanks!
Frank

Maybe you can start here, and see if this helps you. I will search through my handbook of eye movements and see if there is something else in there. The problem with direct gaze contact is, of course, if you don't have eyetracking, it doesn't seem as if the agent is engaging in joint attention, and participant will not see the agent interactive nor responsive. Now, if that doesn't matter to you, you are fine, just make the avatar blink less often, per literature. If you are trying to establish some sort of joint interaction between an avatar and a participant, you will have to give them some type of an object to focus on, and you will have to make the avatar artificially responsive to participant's gaze (tell participant to look at the object B, and wowmiracle, avatar looks there, too).

Also, if your avatar is supposed to be a dominant personality, like a teacher or a tutor or a presenter, it doesn't really matter if there is joint attention or interaction between your participant and an avatar, dominant personalities usually just initiate joint attention and it is up to the participant if they will follow avatar's gaze or not.

You can also make your avatar more or less trusting, depending on the amount of direct gaze it produces.

Just out of curiosity, what are you working on? I might be attacking you with social psychology and you are making a cartoon where everything is prescripted :)

Thank you for the chapter, looks interesting!

My background is in social psychology and so is the nature of this research, so please attack me with that area :)! I'm doing research on how to make a virtual agent more trustworthy. The nature of the 'interaction' with the agent is that participants have to play risky games with the agent (investment game and a route planner game). The whole idea is to measure trust in the agent. In a previous study, I have used mimicry successfully as one strategy to increase trust. However, that agent did not maintain eye contact. Now, I want to test the interaction between eye-contact en mimicry. So do mimicry and eye-contact both add to trust, or is one or another enough to increase trust?

I have found some sources about increasing trust with eye gaze/eye-contact, although it was mostly with humans, and not avatars/agents. So if you know some sources in that area, I would be most interested in them.