July 8, 2026 · Wednesday2026 年 7 月 8 日 · No. 4 NEWSACCESSABOUT

AUTOSIGNAL 车智信号

Human-curated. Expert-annotated. Every signal traced to source. 人工精选 · 专家点评 · 每条信号可溯源

← All signals← 返回全部信号

Research Frontier研究前沿

Guided Action Flow: Q-Guided Inference for Flow-Matching Vision-Language-Action PoliciesGuided Action Flow:Q引导推理用于流匹配视觉语言动作策略

S 1.7 T1 1 sources1 个来源 R7-research
  1. Researchers propose Guided Action Flow, an inference-time framework that uses a learned critic to guide the reverse-time sampling of pretrained flow-matching vision-language-action (VLA) policies without retraining the base model.
  2. On LIBERO manipulation tasks, a single-task critic improves success rates from 68.0% to 82.0% on one seed window (14% gain) and from 82.0% to 86.0% on another (4% gain), while a multi-task critic achieves 10% improvement on validation data (46.0% to 56.0%).
  3. The critic is trained from real success and failure rollouts and conditions on task features from the frozen language pathway, but held-out test set gains are modest at 2.5 percentage points (65.0% to 67.5%), indicating limited cross-task generalization.
  1. 研究人员提出Guided Action Flow框架,在推理时使用学习的批评家引导冻结的流匹配视觉语言动作策略的采样,无需重新训练基础策略。
  2. 在LIBERO操作任务上,单任务批评家在一个种子窗口上从68%提升至82%(提升14%),在另一个种子上从82%提升至86%(提升4%),多任务批评家在验证集上提升10%(从46%至56%)。
  3. 批评家由真实成功和失败轨迹训练,利用冻结语言路径的任务特征调节,但保留测试集的性能提升有限(仅2.5%),反映出跨任务泛化能力受限。