File size: 2,899 Bytes
f1ae877
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
---
license: apache-2.0
language:
- en
task_categories:
- text-generation
tags:
- physics
- rigid-body
- pymunk
- next-frame-prediction
- simulation
- icml-2026
size_categories:
- 1M<n<10M
---

# physics-scenarios-packed

Packed (tar.gz) version of a 2D rigid body physics dataset for training language models on next-frame prediction. **1,000,020 scenes × 200 frames** simulated with Pymunk / Chipmunk2D.

This repo is bandwidth-friendly: each scenario type ships as a single `.tar.gz`. For the unpacked JSONL files see [physics-scenarios-raw](https://huggingface.co/datasets/AlexWortega/physics-scenarios-raw).

## Scale

- **Train**: 900,000 scenes (24 seen scenario types × 37,500 each)
- **Val**: 100,020 scenes (30 scenario types × 3,334 each — includes 6 OOD types)
- **Frames per scene**: 200
- **Format**: JSONL inside tar.gz (1 header line + 200 frame lines per scene file)

## Layout

```
manifest.json                    # full dataset spec
train/
  avalanche.tar.gz               # 24 seen-type archives
  basketball.tar.gz
  ...
val/
  avalanche.tar.gz               # 30 archives (24 seen + 6 OOD)
  pong.tar.gz                    # OOD types — never seen at train time
  bowling.tar.gz
  ...
```

## Scenario types

**Seen (24)** — train + val: avalanche, basketball, billiards, breakout, bridge, chain, conveyor, dominos, explosion, funnel, head_on, jenga, marble_run, orbit, pendulum, pinball, plinko, projectile, pyramid, seesaw, ski_jump, tower, wind, wrecking_ball

**Unseen / OOD (6)** — val only:
- *Simple*: pong, bowling, ramp_roll
- *Complex*: angry_birds, hourglass, newtons_cradle

The 6 unseen types are held out to measure generalization to scenarios with new dynamics or interactions.

## Schema (per scene file)

Line 1 — header:
```json
{"scenario": "billiards", "difficulty": 3, "static_geometry": [...], "constraints": [...], "objects": [...]}
```

Lines 2–201 — frames:
```json
{"frame": 0, "objects": [{"id": 0, "x": 1.234, "y": 5.678, "vx": ..., "vy": ..., "angle": ..., "omega": ...}, ...]}
```

## Categories

The 30 scenario types span 6 categories:
- **Collision** — head_on, billiards, breakout, bowling, pong
- **Stacking** — jenga, tower, pyramid, dominos
- **Ramp** — ski_jump, projectile, ramp_roll, marble_run
- **Constraint** — pendulum, chain, bridge, newtons_cradle, hourglass
- **Minigame** — pinball, plinko, basketball, angry_birds
- **Complex** — avalanche, explosion, wind, wrecking_ball, conveyor, funnel, orbit, seesaw

## Usage

```python
from huggingface_hub import snapshot_download
import tarfile

snapshot_download("AlexWortega/physics-scenarios-packed", repo_type="dataset", local_dir="data")

with tarfile.open("data/train/billiards.tar.gz") as tar:
    tar.extractall("data/train_extracted")
```

## Citation

ICML-2026 submission (in progress). Generation pipeline: `src/physics/scenario_generator.py`.