gymrec 0.1.1

Record and replay gameplay from Gymnasium environments as Hugging Face datasets

🎮 Record and replay gameplay from Gymnasium environments as Hugging Face datasets 📊

gymrec is a Python CLI for collecting gameplay from Gymnasium environments and saving it as replayable Hugging Face datasets. It supports human keyboard capture, built-in agent policies, local dataset storage, Hub uploads, playback verification, and MP4 exports.

It works across Atari through ALE-py, Stable-Retro console environments, and VizDoom. Recordings store observations, actions, rewards, episode metadata, collector provenance, and the gymrec version used to collect the run.

Install

git clone https://github.com/tsilva/gymrec.git
cd gymrec
uv sync
cp .env.example .env

Add your Hugging Face token to .env when you want to upload datasets:

HF_TOKEN=your-api-token

Run the CLI from the repo root with uv run gymrec ... or uv run python main.py ....

Commands

uv run gymrec login                                      # authenticate with Hugging Face Hub
uv run gymrec list_environments                         # list Atari, Stable-Retro, and VizDoom envs

uv run gymrec record BreakoutNoFrameskip-v4             # record human gameplay
uv run gymrec record BreakoutNoFrameskip-v4 --dry-run   # save locally without upload prompt
uv run gymrec record SuperMarioBros-Nes --agent random --headless --episodes 100
uv run gymrec record BreakoutNoFrameskip-v4 --agent breakout --headless --episodes 50
uv run gymrec record BreakoutNoFrameskip-v4 --agent random --headless --episodes 100 --workers 5

uv run gymrec upload BreakoutNoFrameskip-v4             # upload new local episodes to Hub
uv run gymrec playback BreakoutNoFrameskip-v4           # replay recorded actions
uv run gymrec playback BreakoutNoFrameskip-v4 --verify  # compare replay frames against recorded frames

uv run gymrec video BreakoutNoFrameskip-v4              # export all episodes to MP4
uv run gymrec video BreakoutNoFrameskip-v4 --range 3-7  # export a 1-based episode range
uv run gymrec video BreakoutNoFrameskip-v4 --first 5
uv run gymrec video BreakoutNoFrameskip-v4 --last 5

uv run gymrec import_roms ./roms                        # import Stable-Retro ROMs
uv run gymrec minari-export BreakoutNoFrameskip-v4      # export local data to Minari format

Usage

Human recording opens a pygame window. Press Space to start recording, use the environment-specific controls printed in the terminal, press Tab to toggle the overlay, use +/- to adjust FPS, and press Esc to stop.

Agent recording supports human, random, mario, and breakout. --headless is for agent mode only and requires --episodes; --workers runs parallel headless collection and cannot exceed the requested episode count.

Playback uses the local dataset first, then falls back to the Hugging Face Hub dataset repo. Video export requires ffmpeg and writes MP4 files from local data or downloaded Hub data.

Notes

• Requires Python >=3.12,<3.13 and uv.
• Hugging Face uploads use dataset repos named {username}/gymrec__{encoded_env_id} by default.
• Local datasets are stored under ~/.gymrec/datasets by default.
• config.toml controls display scale, FPS defaults, local storage, dataset metadata, and overlay defaults.
• keymappings.toml controls Atari, VizDoom, and Stable-Retro keyboard bindings.
• On Apple Silicon, uv sync installs the committed native Stable-Retro wheel through the project configuration.
• ffmpeg must be available on PATH for video exports.
• minari-export requires Minari; install it with uv sync --extra minari or uv pip install 'minari>=0.5.0'.

Architecture

License

MIT