Astra: General Interactive World Model with Autoregressive Denoising
TLDR
Astra is an interactive general world model using autoregressive denoising for long-term video prediction with action control across diverse real-world tasks.
Reasoning
Strengths: novel autoregressive denoising architecture with action-aware adapter and mixture of action experts for heterogeneous actions; demonstrated improvements over SOTA. Weaknesses: limited details on datasets and quantitative results in abstract; no mention of limitations or failure cases.
Read-first score
Read-first score 76.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.
Field roles
Rank sensitivity
Stability: volatile; rank range: 18.
Keyword Scores
Deep Analysis
Innovations
- Autoregressive denoising architecture for interactive world modeling
- Temporal causal attention to aggregate past observations and support streaming outputs
- Noise-augmented history memory to balance responsiveness and temporal coherence
- Action-aware adapter that injects action signals into the denoising process
- Mixture of action experts for dynamic routing of heterogeneous action modalities
Methodology
Astra employs an autoregressive denoising architecture with temporal causal attention to process past observations and generate streaming future frames. It uses a noise-augmented history memory to prevent over-reliance on past frames, an action-aware adapter for precise action control, and a mixture of action experts to handle diverse action modalities across tasks like autonomous driving and robot grasping.
Key Results
Experiments across multiple datasets demonstrate that Astra achieves improvements in fidelity, long-range prediction, and action alignment over existing state-of-the-art world models.