COPHY: Counterfactual Learning of Physical Dynamics from Visual Input

COPHY is a benchmark for assessing causal physical reasoning in a synthetic 3D environment. It was developed by Fabien Baradel of the Institut National des Sciences Appliquées de Lyon to pose the new problem of learning object mechanics from visual input. The dataset likely contains sequences of mechanical experiments, such as falling towers or colliding objects, for training models to predict outcomes after interventions.

Use Cases

• Benchmarking causal physical reasoning models based on the described synthetic 3D environment.
• Training video prediction models for mechanical systems based on sequences of falling blocks or bouncing balls.
• Developing counterfactual learning algorithms based on the task of predicting outcomes after altering initial conditions.
• Researching latent representation learning for physical properties based on unsupervised, end-to-end training described.

Strengths

• Designed to assess state-of-the-art models for a specific, novel problem of counterfactual physical reasoning.
• Includes a variety of mechanical experiments such as falling towers, bouncing balls, and colliding objects.
• Proposed model is reported to achieve predictions at the level of super human performance.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count, file formats, and data size are unknown, which may limit suitability assessment.
• Last update date is unknown; freshness unverified.

Provenance

Source

Fabien Baradel, Institut National des Sciences Appliquées de Lyon

Collection Method

Developed as a benchmark for a research problem; likely contains synthetic data generated in a 3D environment.

Time Range

Temporal coverage is unknown.

Freshness

Last updated date is unknown.

Geography

Spatial coverage is not applicable for a synthetic benchmark.