Imagined Rollouts Are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure

Jul 16, 2026

Upload by Levi

Long-horizon failure in world models is commonly explained as compounding error without specifying which error structure grows. This work proposes a kinematic-vs-dynamic reframing: world models often imagine kinematically rather than dynamically. The paper operationalizes this view with an imagined Kinematic-Consistency Error (iKCE) per rollout step and a perturbation protocol crossing a physical regime boundary. On DreamerV3 trained for DMC walker-walk, iKCE is orders of magnitude larger than matched real-physics rollouts and stays statistically flat during a friction sweep despite reward collapse, yielding a kinematic-not-dynamic signature beyond the gait period.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

Introduction

Central Claim

Contributions

Evidence

Imagined Rollouts Are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure Finn Rasmus Schfer 1 , Korbinian Moller 1 , Yuan Gao 1 , Christian Oefinger 1 , Sebastian Schmidt2 and Johannes Betz 1 1Autonomous Vehicle Systems Lab Technical University of Munich, Garching b. M¨unchen, Germany [Email: finn.schaefer@tum.de](Email: finn.schaefer@tum.de) 2Data Analytics and Machine Learning Group Technical University of Munich, Garching b. M¨unchen, Germany arXiv :2607 .05966v 1 [ cs .RO] 7 Jul 2026 Abstract—Long-horizon failure in world models is conventionally attributed to compounding error, a generic framing that does not distinguish what kind of error compounds. We propose a kinematic-vs-dynamic reframing: world models tend to imagine kinematically rather than dynamically. We operationalize this asthe imagined Kinematic-Consistency Error, a per-step diagnostic that measures how far a rollout departs from a closed-form kinematic null, paired with a perturbation protocol that tests whether iKCE responds when physical conditions cross a regime boundary. We instantiate the diagnostic on a released DreamerV3 checkpoint trained on DMC walker-walk, where imagined iKCE runs roughly two orders of magnitude above that of matched real-physics rollouts. Across a friction sweep that crosses the gaitcollapse boundary, the model’s iKCE stays statistically flat even as the trained policy’s reward collapses through the same range, providing the kinematic-not-dynamic signature. The diagnostic distinguishes kinematic from dynamic imagination at horizons longer than the embodiment’s gait period. I. Introduction World models have become a load-bearing component of recent embodied AI, serving as latent simulators for planning [3, 4, 5] and as generative environments for selfsupervised learning [6, 7] . A widely-noted failure mode is the deterioration of imagined rollouts over long horizons, conventionally attributed to compounding error [8] . This framing is accurate but underspecified: it does not distinguish what kind of error compounds or which feature dimensions deteriorate. Across four recent Observations in driving VLM/VLA and trajectory-prediction benchmarks indicate that the deterioration carries a specific structural signature that the compoundingerror framing obscures. We adopt the classical mechanics distinction between kinematic and dynamic motion. We define kinematic as motion described purely through position, velocity, and acceleration time series, without invoking the forces or physical constraints that produced it. As dynamic, we define motion that requires those constraints (e.g., mass, friction, contact) to be reproduced correctly. Central Claim Current world models imagine kinematically rather than dynamically: they extrapolate position-velocityacceleration trajectories that are internally consistent with linear kinematic update rules, but inconsistent with the physical constraints that produce real motion. Kinematic fallback is a third, structurally distinct account of long-horizon world-model failure, alongside the two positions that dominate the literature: predictable-representation engineering (the Dreamer line [5]) and error-compounding bounds (MBPO [8]) . Predictable-representation engineering attributes long-horizon reliability to stable, predictable latent representations and pursues it through normalization and balancing techniques; a single algorithm with fixed hyperparameters across 150+ tasks empirically validates the position, yet it is silent on what those representations should contain. Errorcompounding bounds derive an explicit quadratic-in-horizon bound on the gap between model-based and true returns under policy distribution shift, and respond pragmatically by limiting model trust to short branched rollouts from real states. A world model whose latent contains rich kinematic features but no dynamic features satisfies both: stable enough for Dreamerline predictability

FAQs

What new explanation does the paper propose for long-horizon world-model failure?

It reframes the failure as kinematic imagination rather than dynamic constraint reproduction, instead of treating it as generic compounding error.

How is the imagined kinematic-consistency error (iKCE) used?

iKCE is defined as a per-step diagnostic that measures how far a rollout deviates from a closed-form kinematic null, making the claim testable.

What empirical signature supports the kinematic-not-dynamic diagnosis?

Across a friction sweep that crosses a gait-collapse regime boundary, iKCE remains statistically invariant while the trained policy’s reward collapses, indicating regime-invariant kinematic imagination.

Jul 16, 2026

Upload by Levi