Future Dynamic 3D Reconstruction: A 3D World Model with Disentangled Ego-Motion

Abstract

Forecasting the evolution of dynamic environments is crucial for autonomous agents. While generative world models have recently achieved high photorealism in 2D video synthesis by mixing ego-motion and environmental dynamics within the image plane, they exhibit physical inconsistencies, such as morphing or vanishing objects, especially over long time horizons. In this paper, we propose FR3D, a world model that predicts a persistent 3D latent representation for future dynamic 3D reconstruction. Unlike prior works that treat the world as a sequence of image-based features, FR3D explicitly decouples the 3D evolution of the scene from the agent's trajectory, treating the inferred ego-motion as a latent proxy for action. This disentanglement resolves the ambiguities between self-motion and world-motion, ensuring geometric consistency into the future. Furthermore, we introduce a teacher-student distillation strategy that leverages the spatial "common sense" of off-the-shelf foundation models, leading to robust zero-shot generalization. Extensive experiments demonstrate FR3D's strong performance for future dynamic 3D reconstruction from monocular observations across multiple datasets, even 2 seconds into the future. Project page: https://fr3d-wm.github.io.

Figures (5)

• Figure 1: The proposed FR3D is a 3D world model predicting future 3D reconstruction of dynamic scenes that takes monocular images as input. FR3D disentangles the forecasting of the induced ego-camera motion from that of the 3D scene structure. As shown in these future predictions of challenging scenes, FR3D successfully handles dynamic scenes with traffic in both directions (above) and estimates turning events smoothly (below).
• Figure 2: The proposed FR3D takes in input a sequence of images as context (up to time $t_N$), and outputs a unified 3D scene reconstruction with ego camera poses autoregressively for the next timestamps (from $t_{N+1}$ onwards) without accessing the corresponding images. Tokens and state are internal representations of the scene from previous frames, and the model estimates future tokens and decodes them into a 3D reconstruction by leveraging an off-the-shelf foundation model (its encoder, its decoder used to combine the current tokens with the state, and its heads), such as CUT3R wang2025cut3r. The future prediction is performed by two masked transformers: one for ego poses and one for scene geometry. FR3D is trained autoregressively in a teacher-student paradigm by mimicking the token space of the frozen foundation model via a smooth L1 loss.
• Figure 3: Qualitative results on challenging zero-shot dynamic scenes from nuScenes caesar2020nuscenes and KITTI geiger2013kitti.
• Figure 4: Zero-shot prediction of our model on the nuScenes dataset. The example shows a turning scenario.
• Figure 5: Failure cases from two scenes of the Waymo dataset, where our model mixes longitudinal and lateral motion for the dynamic objects. In both cases, the ego vehicle is slowing down approaching an intersection. Context and rollouts are displayed together in the 3D reconstruction. CUT3R's reconstruction on the full sequence is included as reference in the bottom row.