PhysSim — synthetic physics supervision for VLMs

Vision-language physical reasoning using Qwen3-VL with training data from MuJoCo and PhiFlow simulators. This repository is a clean export of code, analysis, evaluation helpers, and the workshop paper sources. The full private project (checkpoints, large logs, local datasets) may live one level up on your machine; refresh this folder with python scripts/build_github_export.py from the parent project root.

Highlights

• Synthetic scenes with simulator-derived labels (no human MCQ annotation for training answers).
• LoRA SFT and optional GRPO on a managed API (Tinker); PhysBench-style evaluation scripts.
• Paper LaTeX under paper/ with bundled figures and sample images.

Quick start

python -m venv .venv
.venv\Scripts\activate          # Windows
# source .venv/bin/activate     # Linux / macOS

pip install -r requirements.txt
pip install "git+https://github.com/huggingface/transformers.git"

See docs/ for training and dataset guides. Do not commit .env or API keys.

Layout

See STRUCTURE.md.

Paper

Build instructions: paper/README.md.

License

MIT — see LICENSE.