Today we hear from Vivian Tan of Beast Inc. Vivian has been creating content for virtual worlds and games for more than 10 years, first designing a successful fashion brand in Second Life, followed by an education-focused adventure game featured in a Stanford University MOOC. She is currently developing Beast Pets, an AI-driven, VR sandbox game and today she will provide her insights around AI + interaction design thus far.
Beast Pets was born from a desire to reinvent the virtual pet genre and create magical pets that have the capacity to grow, learn and develop meaningful relationships with people. We designed Beast Pets as a virtual sandbox environment to simulate a physical space where young dragonlings can live, play and interact with humans.
The intersection of VR and AI gives us an unprecedented opportunity to break down barriers by putting people and virtual characters into the same physical environment. VR’s inherently immersive nature also means that nuances in the AI’s behavior are magnified, and small details can make or break the player’s suspension of disbelief. In this post, we provide the challenges we encountered and important lessons we learned while creating Beast Pets, our fantasy sandbox VR experience. Using case studies from our experience, we will share our processes and solutions that you can apply to creating your own AI characters!
Over the course of building Beast Pets, we learned that one of the most important elements in creating believable AI characters is defining and refining their relationship to the player. This relationship framework serves as our north star, and we structure all of our design and technical decisions around it.
When defining the initial relationship framework, we start by asking three questions:
In our case, we knew we wanted to create friendly pet dragons with the competence level of 2-month-old puppies. We also wanted them to have decision-making capacities, but with the capability to become more obedient over time as their relationship with the player grew. Instead of creating a structured game, we decided to build a VR sandbox where we could nurture this particular relationship framework.
For the refinement stage, we use the following four-step process that allows us to fine-tune the AI pets’ behavior and their social dynamics with the player:
While a sandbox environment is a suitable solution for our use case, it’s certainly not the only way to develop a relationship framework between players and AI characters. This iterative process can serve as a guide for creating compelling AI characters for developers building a VR game, narrative, or any other immersive experience.
When we first started building Beast Pets, we thought the most important and obvious thing to focus on was the dragons’ intelligence. What we overlooked was a series of fundamental aspects to the experience which greatly impact the players' perception and relationship with AI NPCs:
The above variables should complement the target relationship framework. In our case, we learned from our playtesters that an imbalance in these variables could transform a relaxing day in the park into a stressful experience, breaking the friendly framework we wanted to achieve. We had to find the “Goldilocks” combination of size, number and proximity so that players would find our dragons friendly and approachable but not too clingy and overwhelming.
Since the standard humanoid character in 3D games is typically 2-meters tall, we decided to make the first iteration of our dragons half that height. While a 1-meter tall dragon may look small and unintimidating on a screen, it made our stockily-built dragons feel like giant, flying rottweilers in VR. We scaled them down again and playtested until we arrived at the ideal size for a non-threatening dragon, which is roughly that of a house cat.
We learned from our playtesters through trial and error that the ideal number of simultaneous AI characters is highly subjective. While a very small minority of players enjoy being swarmed by little baby dragons, most players enjoy the company of only a few pets. Our solution is to simply let players choose how many dragons they want hanging out at any given time.
Lastly, the dragon’s distance from the player is based on a multitude of comfort factors as well as the dragon’s own personality. Assertive dragons will fly closer while shy dragons will linger behind, but all dragons will give the player a couple feet of personal space unless the player actively draws them closer with toys or treats.
Player perception is as important to create compelling AI characters as the underlying technology. Simple variables like size, number, and proximity can make or break the experience. This is where the relationship framework can provide context for striking the perfect balance. In a horror game with AI enemies, for example, any or all of these variables can dialed up over time to create increasingly intense situations.
Conversely, this doesn’t mean that all friendly AI must be tiny, as long as the scale is consistent with the player’s expectations and relationship with the AI. We don’t intend to keep our dragons small forever, but as they grow up, the other variables will need to be recalibrated to maintain a friendly, comfortable atmosphere.
In addition to maintaining a comfortable amount of personal space between the player and dragons, we also considered the amount of “attention pressure” placed on the players. In Beast Pets, our dragons are drawn to whatever captures their interest, such as food, toys or the person in the room. If a dragon wants to attract the player’s attention, they will fly at the player’s eye level and make eye contact.
When multiple AIs are all staring at you in VR, it can get a bit too intense.
We realized that the best way to ease up on attention pressure is for the dragons to behave more like real puppies -- they’ll hang out near you but they won’t stare you in the eyes constantly. One of our solutions is to periodically distract the dragons. Instead of staring straight at the player all the time, they will glance around for interesting objects or even their playmates, nudge at the balls in the player’s virtual pockets, or look at the player’s hands (especially if there are treats!) We basically had to give the baby dragons some amount of attention deficit to prevent them from becoming too creepy.
Attention pressure can also be applied to AI characters in other genres or contexts. In scenarios where the AI wants the player to take an action, for instance, the attention pressure can be increasingly dialed up to create urgency. Meanwhile, an AI that wants a player to follow it can alternate between high attention pressure (getting the player to notice it) and flighty behavior (getting the player to chase it). Whether or not an AI character’s capable of dialog, attention pressure can be used to convey its intent to the player.
So far, we talked about how the AI character’s appearance and behavior influence players’ perception of their relationship. But in an immersive medium like VR, the range of interactions available to players will also have significant impact on their relationship with the AI. While it’s obvious that we wouldn’t give players a virtual weapon if we didn’t want them to shoot it, how they use their virtual hands is often more subtle. Luckily, this is where developers have a great deal of control in guiding player behaviors through hand gestures.
It didn’t take long for us to learn our lesson: when given a fist in VR, players will punch things. Much to our horror, our early-version playtesters immediately started hitting the little dragons. Because we had designed our dragons to be loved and petted (not brutalized) we had only given them a perma-happy emotional state at the time. This created an absurd feedback loop where players would hit even harder, frustrated that the dragons were still smiling at them after this interaction. Needless to say, we quickly took away fists in the next iteration of Beast Pets.
After learning our lesson the hard way, we revisited our trusty guide: the relationship framework. We started to think of virtual gestures as yet another means of building relationships with AI characters. By removing the ability to form closed fists and limit hostile interactions, we made it easier for players to intuit that their relationship with the dragons was supposed to be a friendly one. We also increased the dragons’ emotional range so that they’ll react appropriately if the players are being too rough, reinforcing the friendly relationship framework.
We hope these lessons learned will help you create amazing AI characters in VR! We look forward to the day when our pet dragons will have new AI friends to play with or even AI foes to battle.
If you’d like to play with our baby dragons, we are looking for more dragon-trainers-in-training to join our community and help us improve Beast Pets! You can find us on Twitter (@beastpets), on Discord, or shoot a message to hello [at] beastpets.com.
About Beast, Inc.
Beast, Inc. is a San Francisco startup that brings adorable AI-driven digital pet companions to life. Their flagship product, Beast Pets, is a family-friendly virtual reality sandbox featuring your new magical best friend: a pet baby dragon. Beast Pets is deployed in over 200 location-based entertainment centers globally and recently launched in Early Access on Steam. Beast, Inc. is also part of the Oculus Start program.