AI Rival Peak Behind the Unique “Interactive Reality TV” Game

In the announcement for the debut of Rival peak This week, Genvid and Pipeworks Studios sent an article from lead software engineer Sol Sliwinski. This should be interesting to all AI programmers.

Rival peak Is a new kind of large-scale interactive live event (MILE) that exists entirely on Facebook’s live stream and allows viewers of the stream to interact with what they are seeing in real time.By increasing the population Of rival peak In a world with 12 AI-driven characters, we created a kind of virtual terrarium by allowing viewers to influence and influence the world individually or collectively.

“The Hunger Games meet SimsIs probably the easiest way to explain Rival peakUsing Genvid’s unique interactive live streaming technology, all viewers are demigods empowered to help or interfere with AI residents Rival peak Both large and small.

AI Rival peak Is based on Pipeworks Studios’ unique system called First Order Regressive Numerical Action Planner (FORNAP), built from scratch over the last five years. FORNAP is based on the previous Automated Planner technology. This is a category that includes Goal-Oriented Action Planning (GOAP). Auto-planners have existed for many years in various ways to drive the movement of artificial “entities”, from video game characters to real-world robots.

In FORNAP development, there were two major issues that had to be resolved. The first was to reduce the search space when trying to find a plan. The second is to add the ability to infer continuous real numbers, as opposed to primary predicates, which are the way classical planners model the world, such as the Stanford Research Institute Problem Solver (STRIPS) system. was.

Image: An example of a complex plan with the goal of warming the AI ​​character

FORNAP was originally created to move non-player characters in more traditional RPG-like games.for Rival peak, FORNAP has been significantly modified and expanded, with some actually removing some of the AI ​​character’s behavior and autonomy for behavior, opening up the decision-making process to the viewer.this means Rival peak It’s like a reality TV show where AI characters are influenced by and react to viewer input.

FORNAP-led characters opportunistically switch goals, plan for the future, weight plans based on personal (and viewer-led) preferences, and make other goals and plans readily available. You can explore when you can’t. Characters determine their most important goals, consider their environmental knowledge and memory, and evaluate thousands of possible ways to solve those goals. FORNAP is creative and unpredictable, but it is also deterministic.

Hierarchized above the planner is the simulation. This includes traditional systems such as pathfinding, game logic, and animation logic. The planner runs in a separate thread and is updated asynchronously with the simulation.

In addition, FORNAP is completely independent of the simulation state, making it very easy to scale up AI to multiple machines.

We have built a custom domain for the domain-independent FORNAPAI engine to create the equivalent of an interactive reality TV show. We also built a customized simulation and AI library. Rival peak. In summary, the high-level AI stack looks like this:

  • simulation: The· Rival peak The AI ​​library sends a request to the agent, which is a Unity game object. This is where the physical things actually happen. Rival peak Simulated environment.
  • AI library: A collection of domain-dependent actions and world states. Rival peak Characters can know what they can do.
  • FORNAP: Domain-independent planner. It can be used in any game in its raw state.

In previous prototypes, actions were measured in seconds, with frequent and often severe failures. This didn’t work for a dramatic, large-scale interactive experience that was pulled out: Rival peak, The character is intended to last daily with an 8 hour stretch.

First, we added a social behavior qualifier so that there are social fabric similarities among the 12 AI characters. In FORNAP’s predecessor, the hungry character didn’t hesitate to eat another AI character. Rival peak There are certain social norms, such as cannibalism being taboo.

Image: Navigation system stress test, especially testing agent behavior in tight spaces

Other extensions Of rival peak AI includes changes to existing goals, needs, actions, and status effects for each character’s actions. AI Rival peak We also needed all-new support for driving score and camera events, and event flagging to capture multiple moments as a “B-roll” for inclusion in the weekly summary show “Rival Speaks.”

In addition, Rival Peak’s AI planner and goal systems have been enhanced to support indirect viewer input via interactive live stream overlays.

Viewers can not only regularly vote for each AI character’s next goal or action, but can also influence the length of the character’s action, such as collecting food and ingredients and creating puzzles.

That’s why the inhabitants Rival peak It can operate almost autonomously, and the constant input of thousands of viewers affects everything from prioritizing tasks and goals to the speed and accuracy of each character’s problem-solving abilities.

Each AI inhabitant is driven by a combination of goals and needs, with many different weights combined to control how the character makes decisions. These weights, which often manifest as personality traits, are considered along with items on hand (inventory), physical and emotional state, etc. to determine how goals are assigned, prioritized, pursued, and achieved. I will.

The twist is to help the viewer shape and drive the storyline through individual collective actions, giving the viewer a myriad of ways and opportunities to change the character’s goals.

To explain the number of different weights and constraints needed to create the right level of autonomy in Of rival peak A dozen AI characters: Early in the previous FORNAP prototype, we were trying to get agents to understand how to meet their hydration needs. The initial setup was a moderately thirsty agent with an empty dining room and puddle.

Because it doesn’t constrain the problem properly The agent planned to pick up the puddle and drink it directly, ignoring the cafeteria... Go back to that one drafting board!

Similarly, if cannibalism was an issue in previous FORNAP iterations, we modeled the agent as a “meat container.” The idea is that meat can be “extracted” to satisfy hunger. When you first connect this As soon as they started to get hungry, we saw agents commit suicide (often in horrific ways) in horror. Again, the problem wasn’t sufficiently limited.

As with cannibalism, the combat that existed in the original FORNAP prototype has been removed. That doesn’t mean that various forms of conflict do not occur. Rival peak, But they are far less final in nature. We focused on replacing combat with various social interactions such as greetings, conversations, jokes, flirting, and seeking help.

Goals that AI characters may try to achieve Rival peak There are many. Broadly speaking, they include nutrition, hydration, warmth, exploration, obstacle removal, food and resource collection, and more. Characters have dozens of possible actions available to achieve their personal or joint goals.

Comprehensive Goal Rival peak To reach the bottom of the “project” – the mysterious entity behind the whole experience. Is this really a quasi-video game you are observing and influencing as a viewer, or is there a higher power of secret with its own motives and methods?

The twelve AI characters are unaware of this meta-narrative, and even if they know it, they are too busy to survive and explore their surroundings, leaving it to the audience to solve the ultimate puzzle.

Each character has a typical personality (“joker”, “hot head”, “flirt”, etc.) that distinguishes them from each other, expressive actions, plot-based status effects, and others available to each. Determine the behavior of. character. And like real people, each of the 12 AI residents has their own goals.

Their goals range from socializing and solving puzzles to protection (self, their home, their friends), rest, healing / recovery, and survival (collecting food, eating food, drinking water). I will. While these goals encourage each character’s behavior, viewers can be affected in many ways. This is especially true for A Bathing Ape (APE).

APE is explained in another essay by my colleague Greg Reinmuth. It’s enough to say that APE is the moment when the audience has the greatest potential to influence the character and drive the story.

By replacing many of the instances normally processed by RNG with “crowdsourcing” decisions made by viewers (collective or subset:) only viewers observing winter are the mystery berries discovered by winter. You can vote on whether you should eat. Mori’), Rival peak It represents a kind of epic social experiment. With all observers exercising divine power, does the audience create a paradise, or apocalyptic wasteland, for the twelve inhabitants? Rival peak.

We have millions of viewers collectively and individually become judges and demigods, influencing stories and characters in a fun, unpredictable, “must watch” kind of way. We set out to create a kind of episodic Lost Meets Survivor Interactive Manga to give.

Did you succeed? For the first time, it really depends on the viewer.

Back to top button