Search papers, labs, and topics across Lattice.
This paper introduces RIFT, a novel approach to world action models (WAMs) that eliminates the need for iterative video rollouts by utilizing learned anticipation tokens to construct a complete future key/value cache in a single pass. The study demonstrates that while traditional methods are sensitive to future values and their positions, RIFT achieves comparable success rates to rollout-based models while significantly reducing action-chunk latency by up to 89.1%. The findings suggest that WAMs can effectively reuse a fixed cache for future conditioning, paving the way for faster and more efficient robotic action generation.
RIFT achieves 98.8% success in action generation while slashing deployment latency by up to 89.1%, challenging the necessity of iterative video rollouts.
World action models (WAMs) condition robot actions on predicted futures, but iterative video rollout increases deployment latency. We ask whether action generation requires the evolving rollout trajectory or only its future representation. Across four WAMs on all 40 LIBERO tasks, paired closed-loop interventions show that masking or reassigning future-cache values changes execution and reduces success, indicating sensitivity to future values and their assigned positions. For Joint and Cosmos-2, however, replaying one fixed final-clean key/value (K/V) cache nearly preserves unmodified execution, with $1.7$ to $1.9$~cm end-effector average displacement error and $97.9\%$ to $98.2\%$ success. This separates cache consumption from production: these models can reuse a fixed cache but still require iterative rollout to construct it. We therefore propose RIFT (\emph{Rollout-free Imagination via Future Tokens}), which uses learned anticipation tokens to construct a complete future K/V cache in one backbone pass while retaining the original future-read interface. On LIBERO, RIFT achieves $98.8\%$ success, close to rollout-based Joint, IDM, and LingBot-VA at $98.4\%$ to $98.6\%$, while reducing action-chunk latency by $68.2\%$ to $89.1\%$. On RoboTwin~2.0, RIFT reaches $92.9/92.6\%$ on clean/randomized scenes, the highest observed among the evaluated methods. These results support rollout-free future conditioning without iterative video generation at deployment.
