Vision-language-action model speeds robot control with stage-aware action refining
StairVLA: Stage-Aware Hierarchical Action Generation for Vision-Language-Action Models
Robotics
Summary
Robots that follow instructions from language and visuals need to decide their moves carefully. The authors noticed that early steps in planning actions use both words and pictures for a rough plan, while later steps focus more on recent visual info to fine-tune moves. They created StairVLA, which divides this process into two parts: a high-level plan made less often and a quick local fix made frequently to adjust actions. This method makes robots work faster while keeping good performance in tasks.
What this means in practice
- •For robotics software engineers: Implement faster and more efficient robot control by separating coarse planning from fine action adjustment using stage-aware denoising.
- •For industrial automation teams: Improve robotic task success rates and responsiveness in assembly or manipulation tasks by integrating hierarchical action generation with frequent local corrections.
Authors
Shangyuan Yuan, Xinda Qi, Yujiang Pu, Wenliang Guo, Xiaobo Tan
Abstract
Vision-language-action (VLA) models increasingly rely on diffusion- or flow-matching-based action heads to generate continuous robot actions. These action heads typically process the denoising trajectory in a largely uniform manner. However, we observe that the conditioning focus naturally shifts across denoising stages: early stages combine language instructions and visual observations to establish a coarse action trajectory, whereas later stages place greater emphasis on current visual observations for action alignment. Based on this insight, we introduce StairVLA, a stage-aware hierarchical action generation framework that uses partially denoised actions as a natural interface between coarse long-horizon action generation and local refinement. A high-level VLA performs early denoising to produce a reusable long-horizon partially denoised action trajectory, while a lightweight refiner operates at a higher frequency to refine local action chunks using the latest observations. This design amortizes expensive high-level VLA computation while preserving frequent closed-loop correction. On LIBERO, our GR00T-style instantiation improves average success from 96.5% to 97.8% while reducing amortized inference latency from 115.0 ms to 44.2 ms per action chunk. More broadly, across two VLA backbones, simulation benchmarks, and real-robot tasks, StairVLA consistently reduces inference cost while maintaining strong task performance.