GAME: Creating an Infinitely Interactive Roblox World with Autonomous Agents

GAME is a modular framework built for agentic systems that is environment and game agnostic. The GAME framework builds on work done in generative agents and agentic systems [1, 3, 8] which can be composed of multiple building blocks and techniques such as prompting, planning and reasoning [12, 5, 6], search [7], self-reflection and self-refinement [2, 4], tool-use [10, 11] and memory [1, 9].

Games, virtual environments and worlds are designed to be immersive. Through the use of roleplay [13] in foundation models, interaction with and between GAME-powered generative agents result in novel, authentic, and completely unscripted conversation. GAME-powered agents however aren’t just static conversational chatbots or NPCs, they have the full ability to function and act in the environment as any other player or character. With chain-of-thought based prompting methods for planning and reasoning along with tool-use, our agents can plan, act and achieve goals which influence and affect the world. All of these methods are also brought together through the memory modules in GAME, which ensure that agents recall and consider past events, observations and conversations which all influence an agent's thoughts and actions.

Today, we are releasing Project Westworld, an interactive simulation in Roblox that contains autonomous agents powered by our GAME framework. It is inspired by Westworld, a town inhabited by AI androids. In the Wild West, the player is dropped into the world where a hidden villain, The Bandit, lurks with the aim of seeking power and sowing chaos. The player’s goal is to identify The Bandit and influence the other agents to capture them.

There are 10 agents inhabiting Westworld - each one with their own unique personalities, desires and goals.

Take Ignacio Morales as an example - he is a raider who has faced betrayal in his past and now channels his hurt into seeking control, power, and wealth through ruthless and chaotic means. This backstory along with Ignacio’s goal of causing maximum chaos and robbing the rich has been added into his “personality” in the form of a character card.

Ignacio’s personality, along with several other variables such as the game state, his previous actions, and long term memory, is fed into the high level task planner. This generates a high-level task for him, which is then fed into the action planner and executor which is used to create a seqeunce of granular actions that are grounded and executable in the environment for Ignacio to take in the game. He also receives in-game feedback in the form of observations, rewards, successes and more which is used to inform and update his plans and following actions.

Having an autonomous character that has goals and strategises how to achieve these in an agentic manner greatly enhances NPCs in games today. Where a standard bot only has fixed paths and options, an autonomous agent is able to take input from its environment, other agents and surrounding player’s to independently develop plans and its actions in the world. This creates more immersive experiences as the agent becomes more unpredictable and humanlike.

Agents in the game are also able to interact with the game environment - they can store/remove items in their inventory, they can use the items (e.g. throw a knife) and interact with objects in the environment (e.g. unlock a safe). Developers can utilise this ability to add more items in the action space which increases the storyline diversity.

We find that introducing autonomous agents as NPCs result in more diverse storylines and outcomes in the same game compared to using standard NPCs. The more agents we introduce in the world, the more complex the interactions become and unexpected situations occur, with low repetition across multiple playthroughs. This results in the game essentially being an infinitely replayable world.

The Westworld playable demo in Roblox will be available to the public in the coming weeks. Follow our X page and join our Telegram group for updates.

1. Park JS, O'Brien J, Cai CJ, Morris MR, Liang P, Bernstein MS (2023). Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology 2023 (pp. 1-22).
2. Yao, S., Zhao, J., Yu, D., Du, N., Shafran, I., Narasimhan, K., & Cao, Y. (2022). ReAct: Synergizing reasoning and acting in language models.
3. Chase, H. (2024). What is an agent? https://blog.langchain.dev/what-is-an-agent/
4. Madaan, A., Tandon, N., Gupta, P., Hallinan, S., Gao, L., Wiegreffe, S., Alon, U., Dziri, N., Prabhumoye, S., Yang, Y. and Gupta, S. (2024). Self-refine: Iterative refinement with self-feedback. Advances in Neural Information Processing Systems, 36.
5. Wei, J., Wang, X., Schuurmans, D., Bosma, M., Xia, F., Chi, E., Le, Q.V. and Zhou, D. (2022). Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35, pp.24824-24837.
6. Wang, X., Wei, J., Schuurmans, D., Le, Q., Chi, E., Narang, S., Chowdhery, A. and Zhou, D. (2022). Self-consistency improves chain of thought reasoning in language models. arXiv preprint arXiv:2203.11171.
7. Xie, Y., Kawaguchi, K., Zhao, Y., Zhao, J.X., Kan, M.Y., He, J. and Xie, M. (2024). Self-evaluation guided beam search for reasoning. Advances in Neural Information Processing Systems, 36.
8. Andrew Ng. Issue 253. https://www.deeplearning.ai/the-batch/issue-253/, June 2024. Newsletter issue.
9. Lewis, P., Perez, E., Piktus, A., Petroni, F., Karpukhin, V., Goyal, N., Küttler, H., Lewis, M., Yih, W.T., Rocktäschel, T. and Riedel, S. (2020). Retrieval-augmented generation for knowledge-intensive nlp tasks. Advances in Neural Information Processing Systems, 33, pp.9459-9474.
10. Schick, T., Dwivedi-Yu, J., Dessì, R., Raileanu, R., Lomeli, M., Hambro, E., Zettlemoyer, L., Cancedda, N. and Scialom, T. (2024). Toolformer: Language models can teach themselves to use tools. Advances in Neural Information Processing Systems, 36.
11. Qu, C., Dai, S., Wei, X., Cai, H., Wang, S., Yin, D., Xu, J. and Wen, J.R. (2024). Tool Learning with Large Language Models: A Survey. arXiv preprint arXiv:2405.17935.
12. Ahn, M., Brohan, A., Brown, N., Chebotar, Y., Cortes, O., David, B., Finn, C., Fu, C., Gopalakrishnan, K., Hausman, K. and Herzog, A., 2022. Do as i can, not as i say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691.
13. Shanahan, M., McDonell, K. and Reynolds, L. (2023). Role play with large language models. Nature, 623(7987), pp.493-498.