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

2026-07-07

AI Summary

The fundamental cause of failure in existing world models during long-horizon prediction remains poorly understood and is often vaguely attributed to error accumulation without distinguishing error types. This work proposes decoupling model imagination into kinematic and dynamic components, introducing a kinematic-dynamic perspective to reframe the failure mechanism. To this end, we develop the imagined Kinematic Consistency Error (iKCE) metric and a friction perturbation protocol to quantitatively diagnose model behavior when crossing physical state boundaries. Experiments on the DMC walker-walk task using DreamerV3 reveal that the model’s imagined trajectories exhibit iKCE values approximately two orders of magnitude higher than those of real trajectories, yet remain stable even when policy performance collapses due to gait failure. This indicates that the model’s imagination is predominantly kinematic-driven, validating the efficacy and insightfulness of our proposed framework.

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 as the 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 gait-collapse 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.

Problem

• world models
• long-horizon failure
• kinematic-consistency
• dynamics
• imagined rollouts

Innovation

• kinematic-consistency error
• world models
• long-horizon prediction
• dynamic imagination
• rollout diagnosis