sarahnator/genesis-physical-interventions

Genesis Physical Intervention Benchmark

This dataset contains controlled Genesis simulations for evaluating physically viable world models in embodied AI settings.

Repository: sarahnator/genesis-physical-interventions Visibility at upload time: private

Scenes

• Ramp-cup-water
• Robotic pour
• Pendulum

Task Types

• Single-rollout outcome prediction
• Scalar physical prediction
• Inverse parameter prediction
• Pairwise counterfactual comparison

Files

data/vlm_entries.jsonl
data/manifest_all.jsonl
scenes/
artifacts/frames/
artifacts/videos/
artifacts/trajectories/

Dataset Size Summary

Number of VLM entries: 111

Entries by scene:

{
  "pendulum": 38,
  "ramp_cup": 31,
  "robotic_pour": 42
}

Entries by query type:

{
  "binary_outcome": 53,
  "inverse_parameter_prediction": 12,
  "pairwise_counterfactual": 15,
  "scalar_prediction": 31
}

Loading Example

from datasets import load_dataset
ds = load_dataset("sarahnator/genesis-physical-interventions", data_files="data/vlm_entries.jsonl", split="train")
print(ds[0])

Limitations

• Labels are simulator-derived, not real-world measurements.
• Some labels use proxies, such as receiver_fraction for robotic pouring.
• This is a research benchmark for physical reasoning, not a complete real-world robotics benchmark.