Experience DOOM powered by Amazon Bedrock.
DOOM-Bedrock leverages VizDoom and the LevDoom benchmark, which offers various difficulty levels based on visual modifications. While LevDoom was originally designed for research in deep reinforcement learning generalization, our project takes a different approach. We harness the power of Amazon Bedrock to process screen data using multimodal Large Language Models (LLMs), determining the next actions in gameplay.
Our focus is on the Seek and Slay module, where the AI agent navigates the map to locate and engage enemies. This unique combination of classic gaming and cutting-edge AI showcases the potential of LLMs in complex, visual decision-making tasks.
- python3.9+
To set up DOOM-Bedrock, follow these steps:
git clone https://github.com/aws-banjo/DOOM-Bedrock
cd DOOM-Bedrock
pip install -r requirements.txt
git clone https://github.com/TTomilin/LevDoom
cd LevDoom
pip install -e .
cd ..To get started with Amazon Bedrock:
- Enable the required models in your Amazon Bedrock console :
- We are using Claude 3.5 Sonnet for this project. For detailed information on available models, consult the Amazon Bedrock User Guide.
- Update the region in the code (line 13 of the main script) to match your AWS region:
# Setup bedrock
bedrock_runtime = boto3.client(
service_name="bedrock-runtime",
region_name="us-west-2", # Replace with your region
)To Start you can run
python doom_berdock.pyLevDoom offers various environments to play in. You can modify the environment by changing the code on line 170:
env = levdoom.make(
"SeekAndSlayLevel1_2-v0"
) # Can replace witgh other LevDoom LevelsAvailable environments:
| Level | Environment | Map |
|---|---|---|
| 0 | SeekAndSlayLevel0-v0 | default |
| 1 | SeekAndSlayLevel1_1-v0 | blue |
| SeekAndSlayLevel1_2-v0 | red | |
| SeekAndSlayLevel1_3-v0 | obstacles | |
| SeekAndSlayLevel1_4-v0 | resized_enemies | |
| SeekAndSlayLevel1_5-v0 | shadows | |
| SeekAndSlayLevel1_6-v0 | mixed_enemies | |
| SeekAndSlayLevel1_7-v0 | invulnerable | |
| 2 | SeekAndSlayLevel2_1-v0 | blue_shadows |
| SeekAndSlayLevel2_2-v0 | obstacles_resized_enemies | |
| SeekAndSlayLevel2_3-v0 | red_mixed_enemies | |
| SeekAndSlayLevel2_4-v0 | invulnerable_blue | |
| SeekAndSlayLevel2_5-v0 | resized_enemies_red | |
| SeekAndSlayLevel2_6-v0 | shadows_obstacles | |
| 3 | SeekAndSlayLevel3_1-v0 | blue_mixed_resized_enemies |
| SeekAndSlayLevel3_2-v0 | red_obstacles_invulnerable | |
| SeekAndSlayLevel3_3-v0 | resized_shadows_invulnerable | |
| 4 | SeekAndSlayLevel4-v0 | complete |
This project is built upon the LevDoom environment. We extend our gratitude to its creators:
@inproceedings{tomilin2022levdoom,
title = {LevDoom: A Benchmark for Generalization on Level Difficulty in Reinforcement Learning},
author = {Tristan Tomilin and Tianhong Dai and Meng Fang and Mykola Pechenizkiy},
booktitle = {In Proceedings of the IEEE Conference on Games},
year = {2022}
}