astra-vision/newtphys_dataset

NewtPhys dataset

This is the dataset used it NewtPhys for benchmarking VLMs ability to sense Newtonian physics. Its primary use is benchmarking and improving physical-awareness in vision models and generative AI.

The dataset provides 7k simulation sequences obtained by augmenting real-world scenes from DL3DV with Google Scanned Objects.
Each sequence comes with dense, low-level pixel-aligned ground truth maps — gravity, collisions, material segmentation, scene flow, optical flow, deformation gradients, instances, amodal maps, etc. — along with high-level annotations about physics — object interactions, velocity, names, etc. —.

The simulation code will be released soon too, along with added videos.

📄 Citation

If you use this codebase or the NewtPhys dataset, please cite:

@inproceedings{cavada2026newtphys,
  title={{NewtPhys}: Do Foundation Models Understand Newtonian Physics?},
  author={Sebastian Cavada and Soumava Paul and Tuan-Hung Vu and Andrei Bursuc and Raoul de Charette},
  year=2026,
  booktitle={arXiv},
  url={https://arxiv.org/abs/2606.03986}
}