ExPhy: A Benchmark for Explicit Physical Property Learning in Multi-Object Trajectory Forecasting
Researchers have created a new benchmark called ExPhy to help evaluate AI models' ability to learn and predict physical properties in multi-object trajectory forecasting. The benchmark includes 24,000 simulated scenes with explicit labels for mass, friction, and restitution, allowing for the evaluation of both trajectory forecasting and physical property estimation. A physics-guided model called PhyODE was also developed, which estimates physical properties from observed traj
Researchers have created a new benchmark called ExPhy to help evaluate AI models' ability to learn and predict physical properties in multi-object trajectory forecasting. The benchmark includes 24,000 simulated scenes with explicit labels for mass, friction, and restitution, allowing for the evaluation of both trajectory forecasting and physical property estimation. A physics-guided model called PhyODE was also developed, which estimates physical properties from observed trajectories and uses them to make predictions. On a long-horizon out-of-distribution test, PhyODE reduced average displacement error (ADE) and final displacement error (FDE) by 33.1% and 31.0%, respectively, compared to the strongest baseline.
---
Why it matters: This matters because it provides a more comprehensive evaluation of AI models' ability to understand physical dynamics, which is essential for applications such as robotics, autonomous vehicles, and smart homes.
Source: https://arxiv.org/abs/2608.20009
This article was originally published at: https://arxiv.org/abs/2608.20009