2026-10-10 · 25 min read约 25 分钟
Where does a VLA decide what to do? Poking at the action interface of three robot policiesVLA 到底在哪里决定动作？拆解三个机器人策略的“动作接口”
A weekend-project-that-got-out-of-hand: linear probes, interchange interventions and causal subspace search on OFT, GR00T and PI heads built on the same Qwen3-VL backbone. Readable is not the same as used, the action is committed at the very last flow step, and the causal interface speaks motor, not geometry.一个越做越大的个人探索：在共享 Qwen3-VL 主干的 OFT、GR00T、PI 三种动作头上做线性探针、交换干预和因果子空间搜索。结论是：可读不等于被使用；动作在最后一步 flow 才被真正“拍板”；因果接口说的是运动语言，而不是几何语言。