New Methods Turn Vision‑Language‑Action Models Into Modular, Verifiable Skill Libraries

Overview

• The two preprints were published Wednesday, September 2, 2026, and offer complementary advances: REFACTOR‑VLA discovers reusable, typed motor programs from action rollouts and EmbodiedSkills adds a fixed runtime interface that checks prerequisites and verifies outcomes.
• REFACTOR‑VLA uses a wake/sleep pipeline that clusters short motor fragments under a Behavioral‑Equivalence Kernel computed from rollouts in a learned latent world model and emits typed lambda‑style programs that a library‑conditioned decoder can use.
• The REFACTOR team reports that adding a supervised contrastive (InfoNCE) loss during world‑model warmup greatly improved clustering quality and produced the first real‑LIBERO task‑language library, while simply increasing world‑model size from 188M to 430M parameters worsened results.
• EmbodiedSkills defines an executable‑skill interface that records planning, execution, verification, and recovery so low‑level VLA policies can be swapped or adapted; task‑adapted policies scored high on RoboTwin and LIBERO but performed poorly on memory‑dependent RMBench tasks.
• Together the papers stress three practical points for embodied agents: ground skill discovery in dynamics, prioritize inductive training signals over raw capacity, and add a runtime execution/verification layer to turn VLA policies into reliable closed‑loop agents.