---
license: cc-by-4.0
task_categories:
  - visual-question-answering
  - video-classification
language:
  - en
tags:
  - physics
  - vision-language
  - synthetic
  - mujoco
  - phiflow
  - simulation
  - physical-reasoning
size_categories:
  - 10K<n<100K
modality:
  - image
  - video
  - text
pretty_name: PhysSim-VLM Dataset
---

# PhysSim-VLM Dataset

**Paper:** *Synthetic Physics as Supervision: Learning Real-World Physical Reasoning in Vision-Language Models*  
**Venue:** AI4Physics Workshop @ ICML 2026  
**Authors:** Swastik R, Natesha B V (IIIT Raichur)

## Dataset Description

PhysSim-VLM is a fully synthetic physics-reasoning dataset for training and evaluating vision-language models (VLMs). It contains **15,000 multi-frame scenes** (train: 12,023 / val: 1,477 / test: 1,500) generated from two physics simulators:

- **MuJoCo** — rigid-body dynamics: time-to-collision (TTC), pile stability, projectile trajectory
- **PhiFlow** — continuum fluid simulation: flow direction, viscosity comparison, fluid level

Each example consists of an 8-frame video rollout of geometric objects (coloured boxes, spheres, cylinders) interacting under physical laws, paired with a free-text question and an answer derived directly from simulator ground-truth state — no human annotation involved.

## Intended Use

- Fine-tuning VLMs on physics-grounded visual reasoning
- Studying synthetic-to-real transfer for physical reasoning
- Probing what physics concepts can be taught via simulator supervision alone

## Dataset Structure

| Split | Size |
|-------|------|
| train | 12,023 |
| val   | 1,477 |
| test  | 1,500 |

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `scene_id` | string | Unique scene identifier |
| `task` | string | Task family (e.g., `ttc`, `stability`, `trajectory`, `fluid_direction`, `fluid_viscosity`, `fluid_level`) |
| `frames_b64` | list[string] | 1–8 video frames encoded as base64 PNG strings |
| `reasoning` | string | Free-text chain-of-thought answer derived from simulator state |
| `config` | dict | Scene configuration (object properties, simulator parameters) |

## Data Generation

Scenes are generated using:
- **MuJoCo 3.x** for rigid-body physics (collision detection, gravity, friction)
- **PhiFlow** for fluid simulation (Navier-Stokes incompressible flow)

Generation scripts are available in the project code repository:  
[https://github.com/Swastikr/PhysSim-VLM](https://github.com/Swastikr/PhysSim-VLM)

## Citation

```bibtex
@inproceedings{swastik2026physsim,
  title     = {Synthetic Physics as Supervision: Learning Real-World Physical Reasoning in Vision-Language Models},
  author    = {Swastik R and Natesha B V},
  booktitle = {AI4Physics Workshop at ICML 2026},
  year      = {2026},
  url       = {https://huggingface.co/datasets/Swastikr/PhysSim-VLM-Dataset}
}
```

## License

[Creative Commons Attribution 4.0 International (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/)