WorldFlow predicts future video and 3D motion jointly with transformers
WorldFlow: Joint Video and 3D Motion Prediction via Large-Scale Pretraining
Computer Vision and Pattern RecognitionArtificial IntelligenceRobotics
Summary
Predicting how objects move in the future is important for robots and AI to understand the physical world. The authors designed WorldFlow, a model that learns from both videos and 3D motion data simultaneously, using a large dataset of real and simulated examples. This approach helps the model better guess where objects will be and how they will move next by combining visual and geometric information. They tested it on various tasks and found it predicts object motions more accurately than other methods.
What this means in practice
- •For robot developers: Improve robot motion planning by predicting object trajectories from combined video and 3D motion cues.
- •For augmented reality engineers: Enhance real-time scene interaction by forecasting future object movements through joint video and 3D prediction.
Authors
Yuhao Lu, Jiongxu Chen, Hexiang Wei, Jiajin Wu, Juelin Li, Mingjian Liang, Fan Yang, Zhuowei Liang, Zihao Liu, Zhennan Wang, Sheng Wang, Lin Shao
Abstract
Predicting future physical states stands as a central challenge in embodied intelligence. Existing video world models leave 3D dynamics implicit, action-based formulations remain tied to specific robot configurations, and 3D motion prediction models use flow-only supervision, providing a limited learning signal for physical dynamics. To circumvent these limitations, we introduce \method{}, a generative world model that jointly predicts future video and query-based 3D trajectories. By parameterizing 3D motion as continuous displacement tokens aligned with the video grid, a unified Mixture-of-Transformers (MoT) backbone jointly denoises both modalities under a rectified-flow objective. This formulation couples explicit geometric dynamics with visual evolution, effectively unlocking the physical commonsense embedded within pretrained video models. To empirically scale and validate this unified architecture, we train \method{} on a heterogeneous mixture of 1.44 million object-level samples spanning simulated and real-world domains. We further introduce WorldFlowBench, comprising 1,678 evaluation samples from 11 data sources across simulated and real-world interactions, including spatial reasoning tasks. Across all three evaluation sets, WorldFlow achieves the lowest average displacement error (ADE) and final displacement error (FDE) among the evaluated learning-based methods. Notably, on the challenging SpatialCog set, WorldFlow achieves an ADE of 0.060 m, a 54.2\% reduction compared with the strongest external baseline. It also reduces ADE and FDE by 55.9\% and 68.5\%, respectively, compared with its flow-only variant. Project page: https://worldflow-rbs.github.io/.