anonsubmiticml2026/PhysicsScenes

PhysicsScenes是一个包含90万个2D刚体物理场景的数据集，这些场景使用Pymunk（基于Chipmunk2D）生成，覆盖6个物理类别中的30种场景类型。数据集分为训练集（900,000个场景，包含24种已见场景类型）、验证集（100,020个场景，包含所有30种场景类型）和OOD验证集（20,004个场景，包含6种保留场景类型）。总大小约为582GB（未压缩），由于体积庞大，此仓库仅提供元数据和1000个场景的样本，完整数据集需请求获取。每个场景以JSONL文件格式存储，包含1个标题行和200个帧行，记录了场景类型、难度、物体位置、速度等信息。场景类别包括碰撞（如台球、保龄球）、堆叠（如金字塔、塔）、斜坡（如斜坡滚动、滑雪跳跃）、约束（如钟摆、链条）、小游戏（如篮球、传送带）和复杂场景（如愤怒的小鸟、沙漏，仅限OOD）。该数据集旨在支持物理动力学建模研究，特别是用于论文“PhysicsLM: Autoregressive Language Modeling of 2D Rigid Body Dynamics”。PhysicsScenes is a dataset of 900K 2D rigid-body physics scenes generated with Pymunk (Chipmunk2D), across 30 scenario types in 6 physical categories. It is split into Train (900,000 scenes, 24 seen scenario types), Val (100,020 scenes, all 30 scenario types), and OOD val (20,004 scenes, 6 held-out scenario types). The total size is approximately 582 GB uncompressed; due to its size, this repository contains metadata and a 1,000-scene sample, with the full dataset available on request. Each scene is stored as a JSONL file with 1 header line and 200 frame lines, capturing scenario type, difficulty, object positions, velocities, and more. Scenario categories include Collision (e.g., billiards, bowling), Stacking (e.g., pyramid, tower), Ramp (e.g., ramp_roll, ski_jump), Constraint (e.g., pendulum, chain), Minigame (e.g., basketball, conveyor), and Complex (e.g., angry_birds, hourglass, OOD only). The dataset supports research in physics dynamics modeling, particularly for the paper "PhysicsLM: Autoregressive Language Modeling of 2D Rigid Body Dynamics".