Conservation and factoring improve physical simulation model reliability
Conservation Buys Stability and Factoring Buys Counterfactuals in Physical World Models
Machine Learning
Summary
Simulated models that try to predict physical systems can fail in two ways: they either drift away from realistic behavior over time or they fail to adapt when physical conditions change. The authors found that using an energy-preserving method helps keep simulations stable for much longer, while explicitly modeling physical connections allows the simulation to correctly respond to changes it hasn’t seen before. These two improvements work independently, meaning one can be fixed without the other. Their findings hold true even when the system's state is inferred from images rather than known exactly.
What this means in practice
- •For roboticists: Improve long-term physical predictions for robots by using energy-conserving simulations to maintain stable behavior and adapt to changing environments.
- •For game developers: Create more reliable physics-based game simulations that remain realistic over long playtime and respond correctly to altered game rules or parameters.
Authors
Yufeng Wang, Parivesh Priye, Lu Wei, Haibin Ling
Abstract
A learned simulator can reproduce its training conditions accurately yet fail in two distinct ways once those conditions change. Over long rollouts, small errors accumulate until the trajectory drifts away from physically plausible behavior; under an intervention on a physical parameter, the model may continue to follow the law seen during training rather than the intervened one. We show that these two failures require different structural remedies. Evolving a learned energy with a symplectic integrator preserves the geometry of the conservative dynamics and keeps rollouts bounded and physically meaningful for up to $100\times$ the training horizon, while equal-capacity predictors, an energy-regularized predictor, and a tuned neural ODE diverge. By contrast, encoding the physical coupling through an explicit linear factorization enables the model to follow a never-seen sign of that coupling, whereas an unrestricted parameterization remains locked to the training law. Crucially, the two mechanisms are separable: removing the structure responsible for long-horizon stability leaves counterfactual transfer intact, while removing the factorized coupling destroys counterfactual transfer without eliminating stability. This double dissociation, established with matched controls that remove or replace one structural component at a time, persists beyond the headline three-body system and remains visible when the physical state must be inferred from pixels rather than provided directly. The result is a concrete design principle for physical world models: long-horizon stability and changed-law generalization arise from distinct structural commitments, and each can be imposed deliberately without requiring the other.