Astra is an interactive general world model designed to predict long-horizon futures from past observations and actions, and is applicable to a variety of general scenarios and different forms of actions.
Use cases: Astra can generate realistic future videos, covering diverse scenarios such as autonomous driving and robotic grasping, and supports precise action interaction, such as camera motion and robot actions. All videos are generated by Astra with a scene image as the first frame.
Main capabilities: Astra proposes an autoregressive denoising architecture, using temporal causal attention to aggregate past observations and support streaming output. It adopts noise-enhanced historical memory to avoid over-reliance on past frames, so as to balance responsiveness and temporal consistency. To achieve precise action control, Astra introduces action-aware adapters that inject action signals directly into the denoising process. In addition, it develops an action mixture-of-experts mechanism that dynamically routes heterogeneous action modalities, enhancing generality in diverse real-world tasks such as exploration, manipulation, and camera control. Astra achieves interactive, consistent, and general long-term video prediction, and supports multiple forms of interaction.
How to use: Astra starts from an initial image, actions, and optional prompts, and generates future videos chunk by chunk. Chunked causal conditioning enforces temporal consistency and faithful action responses. The action mixture-of-experts projects action signals from different modalities into a shared space, enhances them through historical masking, and routes them to modality-specific experts. A dynamic router selects top-k experts, whose outputs are aggregated into a unified embedding and fed into the Flow Transformer, enabling versatile and precise action-conditioned generation.
Applicable scenarios: Astra handles multiple scenarios, including autonomous driving (predicting long-horizon traffic dynamics from control inputs), manipulation (conditioning robot actions on object interaction), and camera control (reflecting viewpoint changes in coherent videos). It also generalizes to scenarios unseen in training, including indoor environments, Minecraft worlds, and animation-style scenes, generating coherent futures that follow camera or navigation commands.
Limitations: The materials do not explicitly list the limitations of Astra. Future work includes releasing complete inference pipelines for additional scenarios such as autonomous driving and robotic manipulation, open-sourcing data processing tools and training code to promote broader reproducibility, and developing a higher-resolution version of Astra by integrating more advanced generative backbones such as Wan 2.2.
