Research map / Probabilistic and scientific modeling / Dynamical systems
CoPhy: Counterfactual Learning of Physical Dynamics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the observed original sequence to infer latent confounders and predict the counterfactual future | CoPhyNet · BlocktowerCF, BallsCF, CollisionCF | MSE on pose: 0.294 vs 0.331; 2.307 vs 3.267; 0.173 vs 0.665 | Helped |
| Estimate stability once for the entire sequence rather than per object at each time step (ablation) | CoPhyNet · BlocktowerCF 3→3 and 3→4 | MSE on 3D pose average over time: 0.285 -> 0.305; 0.478 -> 0.496 | Hurt |
| Replace the GCN with an MLP (ablation) | CoPhyNet · BlocktowerCF 3→3 | MSE on 3D pose average over time: 0.285 -> 0.289 | No clear effect |
| Replace the GCN with an MLP (ablation) | CoPhyNet · BlocktowerCF 3→4 | MSE on 3D pose average over time: 0.478 -> 0.764 | Hurt |
| Predict the whole future in one step rather than autoregressively (ablation) | CoPhyNet · BlocktowerCF 3→3 and 3→4 | MSE on 3D pose average over time: 0.285 -> 0.295; 0.478 -> 0.492 | Hurt |
| Add supervision of confounder estimation (ablation) | CoPhyNet · BlocktowerCF K=4 | MSE on 3D pose average over time: 0.285 -> 0.281 | Helped |
| Add supervision of confounder estimation (ablation) | CoPhyNet · BlocktowerCF K=3 | MSE on 3D pose average over time: 0.478 -> 0.458 | Helped |
| Use ground-truth positions for training rather than estimated positions (ablation) | CoPhyNet · BlocktowerCF 3→3 | MSE on 3D pose average over time: 0.309 -> 0.294 | Helped |
| Use ground-truth positions for training rather than estimated positions (ablation) | CoPhyNet · BlocktowerCF 3→4 | MSE on 3D pose average over time: 0.504 -> 0.482 | Helped |
| Infer confounders from the original sequence without direct confounder labels (ablation) | CoPhyNet · BlocktowerCF K=3 and K=4 | 4-way confounder classification accuracy: Random: 25.0; CoPhyNet: 65.7 and 68.9 | Helped |
| Use the inferred confounder representation together with the intervened frame to predict stability (ablation) | CoPhyNet · BlocktowerCF 3→3 and 3→4 | stability accuracy per block: 71.0 -> 76.8; 69.8 -> 73.8 | Helped |
Closest papers by meaning
- Filtered-CoPhy: Unsupervised Learning of Counterfactual Physics in Pixel Space (ICLR 2022 · similarity 0.90)
- PhysInOne: Visual Physics Learning and Reasoning in One Suite (CVPR 2026 · similarity 0.88)
- ComPhy: Compositional Physical Reasoning of Objects and Events from Videos (ICLR 2022 · similarity 0.88)
- CausalSpatial: A Benchmark for Object-Centric Causal Spatial Reasoning (NeurIPS 2026 · similarity 0.86)
- Apple-$\pi$: Benchmarking Thinking with Video Towards Law-Grounded Physical Intelligence (NeurIPS 2026 · similarity 0.86)
- QUANTIPHY: A Quantitative Benchmark Evaluating Physical Reasoning Abilities of Vision-Language Models (CVPR 2026 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.