Apple-π Dataset Released: Benchmarking Video Models' Physical Reasoning
ziqi_huang_ · x · 2026-08-03
Apple-π is a new law-grounded benchmark designed to evaluate whether video generation models actually reason with physical laws or merely produce plausible-looking motion.
The dataset consists of 400 videos covering 10 canonical classical-mechanics tasks (9 single-law and 1 multi-law). The evaluation is structured into three core stages:
- Perception: Extracting physical quantities or objects from the video.
- Formulation: Deriving the governing physical law.
- Deduction: Generating future states consistent with the law.
The accompanying Orchard dataset is now fully open-sourced on Hugging Face.
More from Research
- CS-JEPA: Decentralized Predictive Architecture for Swarm Robotics — ITMO · 2026-08-03
- ODEWorld: A Continuous-Time Latent World Model via ODE — Dongxiu Liu · 2026-08-03
- N_0-TWAM: First Large-Scale Tactile-Native World-Action Model for Robots — NeoteAIEmbodied · 2026-08-03
- Comma.ai Recaps Hilarious Early Bugs in Self-Driving Models — dosco · 2026-08-03
- Qdrant Open-Sources E-commerce Demo: Millisecond Search Across 5.8M Products — qdrant_engine · 2026-08-03
- Japanese Junior Researchers Launch LENS Community for Computer Vision — HirokatuKataoka · 2026-08-03