Back to blogs
A novel approach to spatial data gathering has unlocked a new, more scalable method of delivering human-like bots to player experiences.
At Modl.ai, our mission is to revolutionize game development by creating smarter, more efficient AI solutions that enhance player experiences and streamline development. A prime example of this mission in action is our work with Riot Games. We created computationally-efficient, human-like bots for tactical shooters, leveraging techniques and technologies that lie at the core of our behavioral AI engine for games.
In a recent scholarly paper, members of the Modl.ai and Riot teams share how we created AI bots for tactical shooter games that mimic human-like behavior efficiently and practically. This approach is not just theoretical. It directly informs the design and functionality of our products, modl:test and modl:play, both of which rely on the same principles to deliver innovative solutions for game testing and gameplay. Let’s explore how these techniques bring life to our bots and, by extension, to games worldwide.
The challenge: Smarter AI, tighter constraints
AI in gaming is no longer about creating bots that are merely competent. The modern player expects bots to act like humans, adding realism and unpredictability to their gameplay experience. Whether it’s for filling multiplayer lobbies or acting as allies or opponents, AI must be engaging and believable. Fall short, and you risk alienating your player base.
Achieving the necessary level of realism presents challenges:
- Computational limits: Most game hardware devotes significant resources to 3D rendering, leaving little room for complex AI.
- Real-time demands: AI needs to react instantly, keeping the gameplay smooth and immersive.
- Behavioral authenticity: Players can quickly spot and exploit robotic behavior. AI must act and make decisions in ways that mimic human unpredictability.
Human-like AI bots in a tactical shooter environment
To test and refine our approach, we used a tactical shooter inspired by VALORANT called Lyra: Ascent. This title served as a testing ground for our bots, allowing us to simulate a realistic multiplayer environment with AI-driven teammates and opponents. Traditional AI models rely on pixel-level visual data analysis, requiring significant computational resources to process game graphics, but we adopted a novel tactic: We used low-resolution ray-cast sensors to capture only essential spatial data such as the locations of enemies, teammates, and objectives, along with higher resolution ray-cast sensors near the player and bot’s focal area. This streamlined approach reduces computational demands, allowing the bots to run seamlessly even on consumer hardware while still capturing the detailed information needed for the bot to act precisely.
With data captured, we set out to achieve human-like behavior using a core strength of our behavioral AI engine for games: imitation learning. By collecting gameplay data from real players, we trained our bots to replicate human actions. The dataset we gathered from Lyra: Ascent included over 48 hours of gameplay, capturing player strategies, movement patterns, and combat decisions. The combined effort resulted in bots that behave not just strategically but convincingly — navigating maps, engaging in combat, and using abilities with human-like finesse.
Real-world applications: modl:test and modl:play
The techniques showcased in Lyra: Ascent aren’t confined to academic research — they’re the foundation of Modl.ai’s core products.
modl:test: Smarter game testing
Game testing is traditionally a labor-intensive process requiring QA teams to identify bugs, validate mechanics, and explore edge cases. With modl:test, our AI-powered bots handle these tasks autonomously, made more effective by computationally efficient AI behaviors.
Using the same sensor-based architecture described above, modl:test bots can navigate game environments to uncover hard-to-reach bugs, stress-test levels, and provide comprehensive reports on crashes, anomalies, and unexpected events.
By automating these processes, developers can focus on creativity while ensuring their games meet the highest standards of quality.
modl:play: Human-like gameplay experiences
When multiplayer games launch, they often struggle to maintain engagement due to low player populations in early stages. modl:play fills this gap by populating games with adaptive bots that mimic human behavior.
These bots learn from player data, adapting their strategies over time to match evolving player styles. Whether as allies or opponents, they provide a dynamic, realistic multiplayer experience, keeping players engaged even before the player base reaches critical mass. To measure success, we employ a number of approaches:
- Behavioral analysis: By comparing metrics such as movement patterns, shooting accuracy, and ability usage, we ensure our bots align with human behaviors.
- Spatial heatmaps: We assess how well bots mimic human navigation strategies by using heatmaps paired with numeric and visual analysis methods in order to visualize and compare with player movement.
- Human testing: In a Turing-test-inspired study, participants watched gameplay clips and answered whether they thought the player was a human or a bot. Remarkably, 30% of bot clips were mistaken for human gameplay, demonstrating the bots’ realism.
One specific bot model stood out, balancing computational efficiency with behavioral authenticity. This model, with only 14.9 million parameters, operates seamlessly on consumer-grade hardware, making it ideal for integration into commercial games.
Why it matters
For developers, human-like bots offer a dual benefit:
- Enhanced player experiences: Adaptive, believable bots from modl:play ensure engaging gameplay, whether filling multiplayer lobbies or acting as opponents in solo modes.
- Better development workflows: Human-like AI bots can also automate testing and debugging, allowing developers to focus on innovation.
For players, bots that behave like humans provide dynamic challenges, unexpected moments, and a sense of realism that elevates the entire game.
The road ahead
As gaming continues to evolve, so do the demands placed on AI. At Modl.ai, we’re pushing the boundaries of what’s possible with AI in gaming. Future developments we are exploring:
- Reinforcement learning: Allowing bots to learn and adapt in real time.
- Transformer models: Enhancing bots’ ability to generalize across diverse scenarios.
- Generative adversarial training: Creating even more lifelike and diverse behaviors.
The work we’ve done with Lyra: Ascent, modl:play, and modl:test, is just the beginning. As we continue to refine our behavioral AI engine for games, we’re committed to delivering tools and experiences that transform the gaming industry — making development faster, smarter, and more creative while ensuring players get the most engaging experiences possible. Contact us to learn more.
Want to see more? Check out this video demo:
