HiddenMass-50K: A Zero-Shot Kinematics Benchmark

This repository contains the official dataset for the STATERA architecture and the HiddenMass Benchmark. It is designed to test if an AI vision model can figure out the hidden physics of an opaque, asymmetric rigid body (specifically, its invisible Center of Mass) just by watching how it moves, spins, and bounces in a video.

Dataset Structure

To protect the integrity of the public benchmark, the dataset is provided in highly optimized HDF5 format for efficient, chunked data loading. It is strictly split into Training data (with physical ground-truth labels) and Blind Testing data (without answers).

1. Data Splits Overview

Split	File Name	Size	Total Sequences	Domain	Ground Truth Included
Train	HiddenMass-50K-Train.hdf5	246 GB	50,000 (40K Active)	Simulation Only	✔ Yes
Ablation	1K-ablation.hdf5	4.89 GB	1,000	Simulation Only	✔ Yes
Test (Public)	HiddenMass-50K-Test-Public.hdf5	49.5 GB	10,063	10,000 Sim, 63 Real	✘ No (For Benchmark)

⚠ Important Note on the 50K Training Set: 10,000 validation sequences inside the 50K HDF5 have been intentionally zeroed-out in-place to prevent train/test leakage. To correctly isolate the 40,000 valid training sequences, you must split the dataset using PyTorch's random_split with a fixed seed of 42. The 1K-ablation.hdf5 is provided as a pre-packaged subset for rapid architectural prototyping without downloading the full 246 GB corpus.

2. Tensor Features & Shapes

To make loading directly into PyTorch seamless, all video sequences are formatted as (C, T, H, W) tensors matching the input requirements of modern Vision Transformers (like V-JEPA).

HiddenMass-50K-Train.hdf5 & 1K-ablation.hdf5

Contains standard input videos and all underlying physical ground-truth.

Feature Name	Shape	Data Type	Description
videos	[N, 16, 3, 384, 384]	uint8	RGB Video frames. Shape represents [Batch, Frames, Channels, Height, Width].
uv_coords	[N, 16, 2]	float32	2D pixel coordinates (x, y) of the true hidden Center of Mass (CoM).
box_center_uv	[N, 16, 2]	float32	2D pixel coordinates (x, y) of the geometric bounding box center.
z_depths	[N, 16, 1]	float32	Absolute metric depth (Z-axis in meters) from the camera to the CoM.
com_magnitudes	[N, 1]	float32	The absolute scalar magnitude of the hidden mass offset.

HiddenMass-50K-Test-Public.hdf5 (Public Evaluation Benchmark)

Contains strictly the input data and domain labels. Ground truth is withheld for benchmarking.

Feature Name	Shape	Data Type	Description
videos	[N, 16, 3, 384, 384]	uint8	RGB Video frames. Real-world footage is dynamically cropped/scaled to precisely match Sim.
domain_label	[N]	string	Text label denoting origin (e.g., "SIMULATION", "REAL_WORLD").

(Note: A small set of real-world videos are provided with their exact coordinates in the main GitHub repository).

Data Generation & Visual Complexity

Standard AI vision models usually suffer from visual-kinematic aliasing (just guessing the visual/geometric center of an object). To force models to actually learn Newtonian physics, our procedurally generated MuJoCo environment actively decouples visual geometry from physical mass:

1. Temporal Constraints: The 16-frame sequence is sampled at 24 FPS (a brief 0.66-second temporal window). Sequences are engineered with high initial angular velocities (> π rad/sec) to guarantee significant off-axis momentum transfer and parabolic apex dynamics.
2. Impact-Centric Windows: You cannot easily calculate hidden mass without observing torque. Sequences often contain collisions with the floor or a ramp, forcing the model to observe the resulting angular momentum and frictional settling.
3. Domain Randomization & Scale: To prevent perspective scale collapse, 53.8% of the dataset applies stochastic camera perturbation. Furthermore, absolute Z-Depth is dynamically scaled from 0.43 m to 14.87 m to force the model to understand true 3D spatial momentum instead of just memorizing 2D pixel trajectories.
4. Hidden Imbalance: The internal asymmetric payload is randomly initialized strictly between 10% and 90% of the bounding geometry's radius. You cannot find the physical center just by looking at the object's surface shape.

Procedural Generation Source Code

The complete MuJoCo simulation code used to generate this dataset is open-source and available in the Official GitHub Repository. You can easily generate your own custom datasets:

1. Navigate to the /sim subfolder in the repository.
2. Open physics_env.py and modify the NUM_EPISODES variable.
3. Run the script. The generator will automatically output the specified number of videos, completely formatted with the exact physical ground-truth tensors provided in the official training data.

Benchmark Submission & Public Leaderboard

Measuring absolute pixel error is a mathematically flawed way to test physics because models can "cheat" by just guessing safely near the geometric center (a phenomenon we identify as Expectation Collapse). To prevent this, we use a custom evaluation suite.

Currently, benchmark evaluation is handled manually via email. To evaluate your model on the HiddenMass Benchmark:

1. Generate predictions for the HiddenMass-50K-Test-Public.hdf5 dataset.
2. Output a standard JSON file containing your predicted (u,v) coordinates for each frame.
3. Email the resulting JSON file to statera@animeshvarma.dev.

What happens next?

• You will receive your results back via email in 2 to 3 business days. This report will include your formal benchmark metrics: Normalized Center of Mass Error (N-CoME), Normalized Kinematic Jitter ($\tilde{J}$), Physics Capture Ratio, and your overall composite Unified HiddenMass Score (HMS).
• Your results email will include an invitation to add your model to the official HiddenMass Public Leaderboard.
• If you choose to opt-in, you will be provided a schema to fill out with your model details (architecture, parameters, organization, etc.) so your model can be publicly ranked against the current State-of-the-Art (SOTA).

(Note: We may build an automated evaluation portal in the future, but currently, this email submission process remains the only official way to benchmark your model).

For full evaluation code, custom metric mathematical formulations, JSON formatting requirements, and the baseline STATERA model architecture, visit the Official GitHub Repository.