This repository is a high-performance GPU batch simulator implementation of OpenAI's Hide and Seek Environment from the paper Emergent Tool Use From Multi-Agent Autocurricula (Baker et al. 2020). The batch simulator is built on the Madrona Engine, and is one of the primary examples used for performance analysis in the SIGGRAPH 2023 paper about the engine: Shacklett et al. 2023. Note that this implementation is not a 1-1 recreation of the original environment. While we have matched many of the high-level details (number of objects, action and observation spaces, etc), lower level details like physics algorithms and procedural generation are implemented differently.

WARNING This repository is not currently under development and does not represent best practices for using the Madrona Engine. The codebase is provided here primarily for the purposes of reproducing performance results from our SIGGRAPH paper. For a clean, well documented, 3D environment written in Madrona with similar functionality, refer to the Madrona Escape Room example project.

First, make sure you have all the dependencies listed here (briefly, recent python and cmake, as well as Xcode or Visual Studio on MacOS or Windows respectively).

Next, fetch the repo (don't forget --recursive!):

git clone --recursive https://github.com/shacklettbp/gpu_hideseek.git
cd gpu_hideseek

Next, for Linux and MacOS: Run cmake and then make to build the simulator:

mkdir build
cd build
cmake ..
make -j # cores to build with
cd ..

Or on Windows, open the cloned repository in Visual Studio and build the project using the integrated cmake functionality.

Now, setup the python components of the repository with pip:

pip install -e . # Add -Cpackages.gpu_hideseek.ext-out-dir=PATH_TO_YOUR_BUILD_DIR on Windows

You can profile the simulator as follows (first, install pytorch):

python scripts/benchmark.py 16000 1920 0 0 1    # Benchmark 16K worlds on the GPU backend
python scripts/cpu_benchmark.py 2000 1920 0 0 1 # Benchmark 2K worlds on the CPU backend

Performance numbers for both backends are expected to be slightly faster than results published in the paper due to engine-level improvements.