ICML-2026/physics-generaliztion

Physics Generalization Dataset

1,000,020 diverse 2D rigid body physics simulation scenes for training and evaluating LLMs on physics prediction tasks.

Overview

This dataset contains procedurally generated physics simulations across 30 distinct scenario types organized in 6 categories. Unlike typical physics datasets that only feature random objects falling in a box, this dataset covers a wide range of physical phenomena: collisions, stacking, ramps, pendulums, constraints, and mini-game-inspired physics.

Each scene is a 200-frame simulation at 1/60s timestep using the Pymunk (Chipmunk2D) physics engine, exported in JSONL format with rich metadata.

Dataset Structure

Splits

Split	Scenes	Scenario Types	Purpose
train	900,000	24 (seen only)	Training
val	100,020	30 (seen + unseen)	Evaluation

Unseen Scenarios (held out from training)

6 scenario types appear only in val, enabling out-of-distribution generalization evaluation:

Difficulty	Scenario	Description
Simple	pong	Ball bouncing between two paddles (zero gravity)
Simple	bowling	Heavy ball rolling toward arranged pins
Simple	ramp_roll	Objects rolling down an inclined plane
Complex	angry_birds	Projectile launched at multi-layer block structure
Complex	hourglass	Objects falling through narrow gap between chambers
Complex	newtons_cradle	Balls suspended by pin joints, momentum transfer

Seen Scenarios (in both train and val)

24 scenario types with 37,500 samples each in train:

Collision & Ballistics: billiards, breakout, explosion, head_on, projectile

Stacking & Structural: bridge, dominos, jenga, pyramid, tower

Ramps & Terrain: funnel, marble_run, plinko, ski_jump (+ unseen ramp_roll)

Pendulums & Constraints: chain, pendulum, seesaw, wrecking_ball (+ unseen newtons_cradle)

Mini-game Physics: basketball, pinball (+ unseen angry_birds, bowling, pong)

Complex & Chaotic: avalanche, conveyor, orbit, wind (+ unseen hourglass)

Data Format

Each scene is a JSONL file (1 header line + 200 frame lines).

Header (line 1)

{
  "type": "scene_header",
  "seed": 1315353,
  "scenario_type": "explosion",
  "scenario_category": "collision",
  "difficulty": 4,
  "description": "Explosion: 25 objects flying outward from center.",
  "object_count": 25,
  "gravity": {"x": 0.0, "y": -981.0},
  "timestep": 0.016666666666666666,
  "static_geometry": [...],
  "constraints": [...],
  "objects": [
    {
      "id": 0, "type": "circle",
      "position": {"x": 401.23, "y": 302.45},
      "material": {"mass": 2.5, "friction": 0.6, "elasticity": 0.7},
      "radius": 15.3
    }
  ]
}

Frame (lines 2-201)

{
  "frame": 1,
  "description": "Frame 1: All objects are in motion.",
  "objects": [
    {
      "id": 0, "type": "circle",
      "position": {"x": 415.67, "y": 318.90},
      "velocity": {"x": 280.5, "y": 320.1},
      "angle": 0.052,
      "angular_velocity": 0.003,
      "material": {"mass": 2.5, "friction": 0.6, "elasticity": 0.7}
    }
  ]
}

Key Features

• 30 scenario types with qualitatively different physics (not just parameter variation)
• Difficulty scaling (1-5) per scenario: controls object count, velocity, structural complexity
• Deterministic generation via seed-based RNG
• Constraints/Joints: PinJoint, PivotJoint for pendulums, seesaws, chains, Newton's cradle
• Custom static geometry: ramps, funnels, peg grids, bumpers, hourglass chambers, basketball hoops
• Rich text descriptions for each scene (useful as LLM context)
• Zero gravity scenarios: billiards, pong, orbit
• Initial velocities: projectiles, explosions, head-on collisions (not just "objects at rest")
• Clean train/unseen split for generalization evaluation

Physics Engine

• Pymunk (Python wrapper for Chipmunk2D)
• Scene: 800×600 pixels
• Fixed timestep: 1/60s
• Elasticity always < 1.0 (energy conservation, no Pymunk instability)
• Threading disabled (determinism)

Generation

Generated using 22 CPU cores in ~29 minutes at ~578 scenes/sec.

python scripts/generate_scenarios_dataset.py --split all --workers 22

File Organization

data_scenarios/
├── manifest.json          # Split config, seen/unseen lists
├── train/
│   ├── avalanche/         # 37,500 scenes
│   ├── basketball/
│   ├── ...                # 24 scenario type directories
│   └── wrecking_ball/
└── val/
    ├── angry_birds/       # 3,334 scenes (UNSEEN)
    ├── avalanche/
    ├── bowling/           # 3,334 scenes (UNSEEN)
    ├── ...                # 30 scenario type directories
    └── wind/

Citation

Part of a research project on training LLMs to predict 2D rigid body physics.