SAMPO:Scale-wise Autoregression with Motion PrOmpt for generative world models

TL;DR

SAMPO tackles the problem of producing temporally consistent, high-fidelity world-model rollouts for action-conditioned planning. It combines scale-wise autoregression with coarse-to-fine intra-frame generation, an asymmetric multi-scale tokenizer, and a trajectory-aware motion prompt to ground dynamics in observed motion. Empirical results show SAMPO achieves strong video prediction quality, improved model-based planning performance, and faster inference, along with zero-shot generalization and favorable scaling behavior. This approach advances interactive, scalable world models for robust planning and control in embodied systems.

Abstract

World models allow agents to simulate the consequences of actions in imagined environments for planning, control, and long-horizon decision-making. However, existing autoregressive world models struggle with visually coherent predictions due to disrupted spatial structure, inefficient decoding, and inadequate motion modeling. In response, we propose \textbf{S}cale-wise \textbf{A}utoregression with \textbf{M}otion \textbf{P}r\textbf{O}mpt (\textbf{SAMPO}), a hybrid framework that combines visual autoregressive modeling for intra-frame generation with causal modeling for next-frame generation. Specifically, SAMPO integrates temporal causal decoding with bidirectional spatial attention, which preserves spatial locality and supports parallel decoding within each scale. This design significantly enhances both temporal consistency and rollout efficiency. To further improve dynamic scene understanding, we devise an asymmetric multi-scale tokenizer that preserves spatial details in observed frames and extracts compact dynamic representations for future frames, optimizing both memory usage and model performance. Additionally, we introduce a trajectory-aware motion prompt module that injects spatiotemporal cues about object and robot trajectories, focusing attention on dynamic regions and improving temporal consistency and physical realism. Extensive experiments show that SAMPO achieves competitive performance in action-conditioned video prediction and model-based control, improving generation quality with 4.4$\times$ faster inference. We also evaluate SAMPO's zero-shot generalization and scaling behavior, demonstrating its ability to generalize to unseen tasks and benefit from larger model sizes.

Figures (15)

• Figure 1: SAMPO Overview. SAMPO is a scale-wise autoregressive world model for video prediction and robotic control. It models temporal dynamics through frame-wise causal generation, while capturing spatial structure via multi-scale tokenization and coarse-to-fine prediction. A trajectory-aware motion prompt further enhances spatiotemporal grounding. SAMPO supports high-fidelity, action-conditioned rollouts for visual planning and model-based reinforcement learning.
• Figure 2: The overall framework of SAMPO. The observed and future frames are discretized by a multi-scale tokenizer to obtain dense and sparse token maps, which are then autoregressively predicted across time, while following a coarse-to-fine decoding order within each frame. Motion prompts extracted from observed frames are injected alongside visual tokens to guide dynamic modeling.
• Figure 3: Motion Prompt generation.
• Figure 4: Qualitative comparison of video prediction. We compare SAMPO with iVideoGPT on the 1X World Model and RoboNet datasets under action-conditioned settings. SAMPO shows a clear advantage in modeling complex backgrounds and capturing dynamic object interactions over time.
• Figure 5: Model-based RL with SAMPO. We report the success rate (in %) across 6 tasks, along with the average success rate across all tasks and the 95% confidence interval agarwal2021deep calculated from 5 random seeds. All models are pre-trained with world models seo2022reinforcement.