Mixture of Horizons in Action Chunking

Vision-language-action models exhibit an inherent trade-off in action chunk length (``horizon''): longer horizons improve global foresight but degrade fine-grained local control, while shorter ones yield the opposite. To mitigate the trade-off, we propose a $\textbf{mixture of horizons (MoH)}$ strategy. In brief, MoH rearranges the action chunk into several segments with different horizons, processes them in parallel with a shared action transformer, and fuses outputs with a light linear gate. It offers three appealing benefits. i) Long-term foresight and short-term precision are jointly exploited within a single model. ii) MoH is plug-and-play for full-attention action modules with minimal training or inference overhead. iii) MoH enables dynamic inference with adaptive horizons, which selects stable actions through cross-horizon consensus, achieving 2.5$\times$ higher throughput than baselines while preserving superior performance. Extensive experiments over flow-based and one-step regression policies demonstrate that MoH yields consistent and significant gains on both simulations and real-world tasks.

(1) Problem: Robots powered by Vision-Language-Action models (VLAs) now learn to perform diverse manipulation tasks like folding clothes by iteratively predicting and executing short sequences of future movements, called "action chunks." But how far ahead the robot should plan presents a dilemma: planning many steps ahead helps it stay on course toward distant goals, while planning fewer steps allows more precise, careful motion. Existing systems must pick one fixed planning length, sacrificing either foresight or precision. (2) Solution: We propose Mixture of Horizons (MoH), a simple strategy that lets the robot plan at several lengths simultaneously and intelligently blend them, like a driver who watches both the road ahead and the steering wheel at once. A tiny gating layer learns when to rely on far-sighted versus near-sighted predictions. We also use agreement across these plans to distinguish how confidently the robot act and realize dynamic inference. (3) Impact: MoH plugs into existing VLAs with negligible cost, consistently improves performance across simulated and real-world tasks, and enables 2.5× usage of predicted actions with dynamic inference while keeping comparable performance. MoH is a step toward more capable, efficient embodied AI.