When robots fail at long-horizon tasks, the bottleneck is usually not "can't do the action" but "did it without knowing whether it worked." Vision-language-action (VLA) models map visual observations and language instructions directly to robot actions, yet an action prediction by itself does not guarantee that the proposed operation is valid in the current physical state, or that its outcome will ever be checked. A paper uploaded to arXiv on Sep 1 (2609.01281) offers an engineering answer: EmbodiedSkills, a unified framework that orchestrates, trains, and deploys VLA agents within a single agent loop.

Core design: every skill decision is an execution proposal

The most interesting choice here is how the framework treats model decisions: every skill decision is merely a proposal. The runtime checks its prerequisites before execution — whether the required observations, plan, and backend are ready — and verifies the outcome afterward using fresh observations. The whole process runs as a six-stage loop: observation, planning, preflight, bounded execution, verification, and recovery. A shared executable-skill interface connects high-level skill selection, bounded low-level VLA execution, and post-action verification within one loop.

Because this interface stays fixed, low-level VLA policies can be replaced or adapted without touching the agent loop. The interface also records planning, execution, verification, and recovery events as structured trajectories — supervision signals for individual components, and a basis for optional online adaptation when interactive feedback is available. The paper lists 17 authors and the code is open-sourced on GitHub (DCDmllm/EmbodiedSkills).

The numbers: verification alone is worth 38 points

The experiments instantiate the framework with Qwen3-VL and OpenPI/pi0.5 on RoboTwin 2.0 and LIBERO. Task-adapted low-level VLA policies reach an 86.20% average success rate across 50 RoboTwin 2.0 tasks (100 episodes per task; the pi0.5 reference baseline is 82.74%) and 97.40% across the four LIBERO suites (OpenPI reference: 96.85%).

The ablations matter more: in a controlled comparison over the same 50 tasks and 5,000 episodes, the full AgentLoop reaches 86.20%; removing intermediate verification drops it to 48.2%; replacing semantic subtasks with the full task instruction yields 34.4%; limiting every subtask to a single action chunk leaves just 19.5%. In other words, the "check after execution" step alone is worth 38 percentage points. On four memory-dependent RMBench tasks, the same task-adapted approach averages only 12.5% — the paper does not hide it: memory remains the weak spot.

So what

This work pulls embodied AI back from "is the single-step prediction accurate" to "is the closed loop trustworthy." The gap between 86.2% and 48.2% shows that for long-horizon physical tasks, verification and recovery are not decoration but load-bearing walls. For teams building robot agents, the open-sourced framework's trainable and inspectable agent layer is an engineering reference worth dissecting — turn "did it act" into "did it act correctly" first, then talk about generalization.

Paper and code: https://arxiv.org/abs/2609.01281