文章背景与核心概要
理解物体动力学不仅需要预测未来的运动轨迹,还要求模型具备捕捉支配运动的底层物理属性的能力。然而,当前的基准测试往往忽视了这些显式的物理属性。为了填补这一空白,本文推出了 ExPhy,这是一个针对多物体轨迹预测的综合基准测试,包含 24,000 个模拟场景,并带有关于质量、摩擦力和恢复系数的显式真实标签(ground-truth labels)。
该研究同时引入了 PhyODE,这是一种物理引导的模型,它利用显式属性接口从观测到的轨迹中估计物理参数,从而实现更准确的可微分未来推演(rollouts)。实验结果表明,PhyODE 显著优于现有的基准模型,特别是在长视距、分布外(OOD)的设置下。值得注意的是,该研究强调:轨迹预测的高准确性并不能保证对底层物理属性的正确恢复。
ExPhy: A Benchmark for Explicit Physical Property Learning in Multi-Object Trajectory Forecasting
Authors: Rui Wang, Yeteng Wu, Xianlin Zhang, Mengshi Qi
Date: August 20, 2026 (v2: August 25, 2026)
Subject: Artificial Intelligence (cs.AI)
DOI: 10.48550/arXiv.2608.20009
摘要
理解物体动力学不仅需要预测未来的轨迹,还需要模型捕捉支配运动的底层物理属性的能力。当前的基准测试往往忽视了这些显式的物理属性。ExPhy 通过提供一个用于多物体轨迹预测的全面基准测试填补了这一空白,其中包含 24,000 个模拟场景,并带有质量、摩擦力和恢复系数的显式标签。
该研究引入了 PhyODE,这是一种物理引导的模型,它利用显式属性接口从观测轨迹中估计物理参数,从而实现更准确的可微分未来推演。实验结果表明,PhyODE 显著优于现有的基准模型,特别是在长视距分布外 (OOD) 设置中。值得注意的是,研究强调,轨迹预测的高准确性并不能保证正确恢复底层的物理属性。
Understanding object dynamics requires more than just predicting future trajectories; it necessitates a model's ability to capture the underlying physical properties governing motion. Current benchmarks often overlook these explicit physical properties. ExPhy addresses this gap by providing a comprehensive benchmark for multi-object trajectory forecasting, featuring 24,000 simulated scenes with explicit labels for mass, friction, and restitution.
The study introduces PhyODE, a physics-guided model that utilizes an explicit property interface to estimate physical parameters from observed trajectories, enabling more accurate differentiable future rollouts. Experimental results demonstrate that PhyODE significantly outperforms existing baselines, particularly in long-horizon out-of-distribution (OOD) settings. Notably, the research highlights that high accuracy in trajectory forecasting does not guarantee a correct recovery of the underlying physical properties.
ExPhy 的核心特征
- 数据集: 24,000 个模拟物理场景。
- 显式标签: 包含质量、摩擦力和恢复系数的真实数据。
- 评估划分:
- ID(分布内): 标准评估。
- OOD-Parameter(属性分布外): 针对物理参数的分布外测试。
- OOD-Initial(初始状态分布外): 针对初始状态的分布外测试。
- 迁移能力: 包含在 ComPhy 基准测试上的零样本(zero-shot)评估,以检验跨基准测试的性能。
Key Features of ExPhy
- Dataset: 24,000 simulated physical scenes.
- Explicit Labels: Includes ground-truth data for mass, friction, and restitution.
- Evaluation Splits:
- ID (In-Distribution): Standard evaluation.
- OOD-Parameter: Out-of-distribution testing over physical parameters.
- OOD-Initial: Out-of-distribution testing over initial states.
- Transferability: Includes zero-shot evaluation on the ComPhy benchmark to assess cross-benchmark performance.
性能亮点
- PhyODE 模型: 一种物理引导的架构,通过估计属性来为未来的运动预测提供信息。
- 结果: 在长视距的
OOD-Initial设置下,与最强基准模型相比,PhyODE 将 ADE(平均位移误差)降低了 33.1%,将 FDE(最终位移误差)降低了 31.0%。
Performance Highlights
- PhyODE Model: A physics-guided architecture that estimates properties to inform future motion.
- Results: In the long-horizon
OOD-Initialsetting, PhyODE achieved a 33.1% reduction in ADE (Average Displacement Error) and a 31.0% reduction in FDE (Final Displacement Error) compared to the strongest baseline.
资源
- 代码与数据: GitHub 仓库
- 完整论文: 查看 PDF
- 许可协议: 知识共享署名 4.0 国际版
Resources
- Code & Data: GitHub Repository
- Full Paper: View PDF
- License: Creative Commons Attribution 4.0 International
