Latent Dynamics · 2026-10-22

Causal world models under physical parameter shifts

Chu Lin, Yi Sun

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

The authors study how causal world-model design affects prediction under physical parameter shifts. In a synthetic shortcut benchmark, a frozen intervention-derived dependency mask reduces mean OOD RMSE from 1.769 to 0.697 (near an oracle 0.694). In CausalWorld link-mass extrapolation, a mass-affine “known-PD” mechanism turns out-of-range masses into consistent predictive gains (median joint R2 0.609→0.727), while other predictors degrade when given out-of-range masses.

Why it matters

The paper argues that both input dependency selection and the mechanism that applies parameter values matter for extrapolating causal physical parameters. It shows that out-of-range parameters can help or hurt depending on how they enter the dynamics model, and that structured mass-affine mechanisms can make actual mass information consistently beneficial under distribution shift.

Method

Study two causal-design components for world models: fixed input dependency constraints (synthetic intervention-derived masks) and dynamics mechanisms for physical parameter effects (mass-affine dynamics).
Create finger-local mass-affine robot world models where link masses enter inertia and dynamics-bias via explicit affine coefficients; actions go through either a learned torque network or a known PD controller; residuals model contact-related effects.
Compare to mass-information controls (remove mass inputs or fix mass inputs at training means) and evaluate both absolute prediction and the predictive benefit of providing actual out-of-range masses. (Also compare a shared 100-step forecast against a recurrent robotic world model.)

Limitation

The paper’s robot evaluation assumes available link masses and a known low-level controller for the PD variant, and covers forecasts of 0.4–0.6 s. It notes that longer deployments, uncertain parameter estimates, and downstream control remain to be evaluated.

Abstract (from arXiv)

World models predict how embodied systems evolve under actions, but their accuracy can degrade when physical conditions move outside the training distribution. We study two components of causal modeling: which inputs a predictor may use, and how its mechanism computes their effects. In a synthetic benchmark with shortcut features, a frozen dependency mask derived from paired interventions reduces mean out-of-distribution (OOD) RMSE from 1.769 for a fully connected predictor to 0.697, close to the generator-oracle mask (0.694). For robot link-mass extrapolation in CausalWorld, with disjoint training and test mass ranges, we develop finger-local world models in which link masses enter inertia and dynamics-bias terms through explicit affine coefficients, actions act through either a learned torque network or the known PD controller, and learned residuals capture contact-related effects. Compared with a separately trained control whose mass inputs are fixed at training means, the known-PD model turns out-of-range masses into a consistent predictive gain: median time-averaged joint R2 over 150-step rollouts rises from 0.609 to 0.727, with improvement in all ten task-seed runs (exact two-sided sign test, p = 0.002). In a shared 100-step forecast, the learned-torque model is competitive with the strongest recurrent robotic world model (RWM) configuration (median joint R2 of 0.755 - 0.768 versus 0.752), whereas supplying out-of-range masses sharply degrades the MLP and recurrent predictors. These results indicate that how a physical parameter enters the dynamics mechanism determines whether its out-of-range value helps or harms prediction.

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-21 · Indexed by Pub-AI: 2026-10-22 UTC

preprint