Learning to Move Like Professional Counter-Strike Players
By: David Durst, Feng Xie, Vishnu Sarukkai, Brennan Shacklett, Iuri Frosio, Chen Tessler, Joohwan Kim, Carly Taylor, Gilbert Bernstein, Sanjiban Choudhury, Pat Hanrahan, Kayvon Fatahalian
[Stanford University & Activision Blizzard & NVIDIA] (2024)
Link to the study: https://arxiv.org/abs/2408.13934
In the high-paced world of multiplayer, first-person shooter games like Counter-Strike: Global Offensive (CS), the way players move and coordinate with their teammates is crucial for success. However, designing movement strategies for every possible scenario in the game is incredibly challenging due to the complex nature of team coordination and the variety of situations players might encounter.
This study explores a data-driven approach to replicate the movement patterns of professional CS players using artificial intelligence. The researchers collected data from 123 hours of professional gameplay and used it to train an advanced AI model, based on transformer architecture, that can mimic human-like movement during a "Retakes" round in CS.
One of the most impressive aspects of this model is its efficiency—it can make movement decisions for all players in less than 0.5 milliseconds per game step on a single CPU core, making it fast enough to be used in real commercial games.
Human testers evaluated the AI's performance and found it to be significantly more human-like than existing bots and even expert-scripted movement controllers. In tests where bots controlled by the AI played against each other, the model showed the ability to perform basic teamwork, make fewer common mistakes, and produce gameplay patterns similar to those of professional players, including where and when players are eliminated during the game.
This research highlights the potential for AI to enhance the realism and sophistication of non-player characters in video games, making them behave more like human players and improving the overall gaming experience.
Info and photos were copied from a socmed post which I can't find anymore since the app refreshed. Credit goes to them and I wanted to share it here. I have not read the entirety of the study.
4 Comments
Info and photos were copied from a socmed post which I can’t find anymore since the app refreshed. Credit goes to them and I wanted to share it here. I have not read the entirety of the study.
hey look a million authors to curb any difficulty at all in publishing, congrats on your field
[Link to pdf](https://www.arxiv.org/pdf/2408.13934)
If you don’t have time to read, here is a summary.
Key Conclusions:
MLMOVE: A new, computationally efficient, transformer-based bot for CS:GO Retakes has been developed. It generates human-like team movement for both teams within the performance constraints of commercial game servers.
Human-like movement: Human evaluators rated MLMOVE’s behavior as significantly more human-like compared to both commercially available bots and expert-crafted rule-based bots.
CSKNOW dataset: A novel 123-hour dataset of professional CS:GO gameplay traces was created, enabling the training of MLMOVE.
Quantitative metrics: New metrics were defined to assess how well a bot’s movement emulates human team-based positioning, and these metrics correlated with human evaluator assessments.
Teamwork and mistakes: MLMOVE demonstrates simple forms of teamwork (like flanking and spreading) and makes fewer common movement mistakes than other bots, aligning its behavior closer to human players.
Movement and outcomes: MLMOVE yields movement distributions, player lifetimes, and kill locations similar to those seen in professional CS:GO matches, further supporting its human-like movement.
Generalization: While focused on CS:GO, the MLMOVE approach could potentially be generalized to other multiplayer FPS games with the development of appropriate datasets and game-specific adaptations.
Very interesting, there is link to [project page](https://davidbdurst.com/mlmove/) and [code](https://github.com/David-Durst/csknow)