Select-to-Act: Hierarchical Reinforcement Learning via Adaptive Language Guidance

Recent advancements in Reinforcement Learning address sample efficiency and trajectory optimization. The Hierarchical Reinforcement Learning with Language Instructions (HRLLI) framework improves efficiency by utilizing adaptive natural-language guidance, overcoming the limitations of single conditioning inputs in complex environments. Simultaneously, Curvature-Adaptive Consistency Flow Matching (CACFM) optimizes consistency distillation for diffusion models. CACFM treats distillation as a dynamic decision process to resolve optimization bottlenecks at boundary stages, enhancing the inference speed and accuracy of autonomous trajectories.