Search papers, labs, and topics across Lattice.
This paper introduces RISE, an adaptive imagination framework for World Action Models (WAMs) that optimizes planning by dynamically allocating imagination resources based on expected benefits. By employing a Latent Evaluator to assess risks and a Rollout Gate to balance computation costs, RISE selectively decides when to continue or halt rollouts, enhancing efficiency. The authors validate their approach using the CounterDrive dataset, demonstrating that RISE outperforms existing methods in planning performance while minimizing unnecessary computations across various WAM architectures.
RISE not only boosts planning efficiency but also redefines how imagination budgets can be adaptively allocated in complex environments.
World Action Models (WAMs) improve planning by incorporating future world evolution into action generation, yet existing methods allocate a fixed imagination budget to every scene. We propose RISE (\textbf{R}efining \textbf{I}magination through \textbf{SE}lective Rollout), a system-level adaptive imagination framework that makes sequential \textsc{Roll}/\textsc{Stop} decisions according to the expected planning benefit of continued rollout. At each step, a Latent Evaluator estimates the risk revealed by the current prefix and how much planning could improve if imagination continues, while a Rollout Gate weighs this expected benefit against additional computation cost. Since factual driving logs expose only one realized future, we further construct \textbf{CounterDrive}, a counterfactual dataset with diverse outcomes and risk levels, to enrich future dynamics and provide localized risk supervision. Each retained sample undergoes expert verification and annotation of trajectory validity, incident onset, and causal category, providing a reusable resource for safety-critical world-modeling research. Experiments on NAVSIM and nuScenes show that RISE achieves the best overall planning performance while reducing unnecessary rollout, with additional transfer results supporting its plug-in generality across WAM architectures.
