awwkl/zwm-intuitive-physics

ZWM intuitive-physics eval dataset

A short-timescale physical-reasoning benchmark used to evaluate world models in the ZWM repo. It contains tabletop hand-object interactions across five categories of physical reasoning:

1. Object cohesion — moving part of an object moves the whole object.
2. Support (top object moves) — removing the support drops the supported object.
3. Support (bottom object moves) — moving the bottom of a stack carries the top with it.
4. Force transfer — pushing object A into B makes B move.
5. Force separation — moving A does not affect a spatially separated B.

The benchmark contains 100 image pairs (5 categories × 20 pairs). For each pair the model predicts a target frame (frame_03) conditioned on a context frame (frame_02) plus a small set of revealed patches around a grounding point. Each pair is evaluated under 8 random mask configurations, yielding 5 × 20 × 8 = 800 evaluations per model.

Accuracy is the proportion of examples for which the model's prediction is closer to the ground-truth target (frame_03) than to the context frame (frame_02), measured in MSE and LPIPS.

Layout

annotations.csv                                       # 100 rows: category, video_id, factual_x/y, counterfactual_x/y, keyframes
keyframes/<category>/<video_id>/frame_02.png          # 512x512 RGB context frame
keyframes/<category>/<video_id>/frame_03.png          # 512x512 RGB target frame
segment_masks/<category>/<video_id>/frame<N>_<primary|secondary|overall>_mask.npy
                                                      # (1, 512, 512) binary masks (secondary may be absent)

Usage with the ZWM repo

huggingface-cli download awwkl/zwm-intuitive-physics --repo-type dataset --local-dir data/evals/intuitive_physics/
bash scripts/eval/intuitive_physics/eval_intuitive_physics.sh
bash scripts/eval/intuitive_physics/grade_intuitive_physics.sh

Source

Derived from the ObjectReasoning benchmark in the counterfactual-benchmark repo (internal Stanford NeuroAI).