Mixture of Horizons in Action Chunking

Vision-language-action (VLA) models have shown remarkable capabilities in robotic manipulation, but their performance is sensitive to the $\mathbf{\text{action}}\mathbf{\text{chunk}}\mathbf{\text{length}}$ used during training, termed $\mathbf{\text{horizon}}$. Our empirical study reveals an inherent trade-off: longer horizons provide stronger global foresight but degrade fine-grained accuracy, while shorter ones sharpen local control yet struggle on long-term tasks, implying fixed choice of single horizons being suboptimal. To mitigate the trade-off, we propose a $\mathbf{\text{mixture}}\mathbf{\text{of}}\mathbf{\text{horizons}}\mathbf{\text{(MoH)}}$ strategy. 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 has three appealing benefits. 1) MoH exploits long-term foresight and short-term precision jointly within a single model, improving both performance and generalizability to complex tasks. 2) MoH is plug-and-play for full-attention action modules with minimal training or inference overhead. 3) 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 policies ${\pi }_0$, ${\pi }_{0.5}$, and one-step regression policy ${\pi }_{\text{reg}}$ demonstrate that MoH yields consistent and significant gains on both simulations and real-world tasks. Notably, under mixed-task setting, ${\pi }_{0.5}$ with MoH reaches a new state-of-the-art with 99$\mathrm{\%}$ average success rate on LIBERO after only $30k$ training iterations. Project page: https://timsty1.github.io/moh/