CF-VLA: Efficient Coarse-to-Fine Action Generation for Vision-Language-Action Policies

Summary

CF-VLA 用 coarse-to-fine 两阶段重构 action generation：coarse stage 构建 action-aware starting point，fine stage 单步修正残余误差。解决 flow-based VLA 的 efficiency-quality trade-off 问题。

Problem & Motivation

Flow-based VLA 问题：

• Multi-step inference 从 Gaussian noise 恢复 action structure 效率低
• Real-time constraints 下 efficiency-quality trade-off 差

Method

核心设计：

1. Coarse stage: 学习 conditional posterior over endpoint velocity，构建 structured initialization
2. Fine stage: 单步 refinement
3. Stepwise training strategy: 先学 controlled coarse predictor，再 joint optimization

优势: NFE=2 优于现有方法，NFE=10 匹配 π₀.₅ baseline

Key Results

• CALVIN 和 LIBERO 验证
• Action sampling latency -75.4%
• Real-robot success rate 83.0%（超过 MIP 19.5pt，π₀.₅ 4.0pt）
• Score 6

Strengths & Weaknesses

亮点：

• Coarse-to-fine 设计合理
• 83.0% real-robot success rate 数字亮眼
• -75.4% latency 显著

局限：

• 与 World Model 关联：这是 action generation efficiency，而非环境建模
• Flow-based VLA 特定优化

Mind Map

mindmap
  root((CF-VLA))
    Problem
      Flow-based VLA 效率低
      Real-time constraint
    Method
      Coarse stage (structured init)
      Fine stage (single-step)
      Stepwise training
    Results
      83.0% success rate
      -75.4% latency

Notes

[基于 arXiv abstract]

VLA action generation efficiency 优化。83.0% real-robot success rate 是有价值的数据点。与 World Model 关联较弱。