Mixture of Horizons in Action Chunking
Researchers have developed a "mixture of horizons" (MoH) strategy to improve the performance of vision-language-action (VLA) models in robotics. This approach addresses the trade-off between long-term foresight and fine-grained accuracy inherent in fixed action chunk lengths. By processing action segments with different horizons in parallel and fusing their outputs, MoH enhances both performance and generalizability across complex tasks. The method is plug-and-play with minimal overhead and enables adaptive inference for higher throughput. AI
IMPACT Enhances robotic manipulation by enabling models to balance long-term planning with fine-grained control.