EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents

EmbodiedSkills. A Unified Framework for Orchestrating, Training, and Deploying VLA Agents. By Wei Wang and sixteen others, from Zhejiang University, the Nanjing University of Aeronautics and Astronautics, and four other institutions. Published on the first of September. Its code repository is worked out: about sixty-five source files, with a readme.

Vision-language-action models, VLA models for short, map what a robot sees and the instruction it is given directly to actions. For long-horizon tasks that is not enough. An agent has to coordinate perception, planning, execution, checking its own progress, and recovering when something goes wrong, all while the physical state keeps changing underneath it. A predicted action, or a skill decision generated by a model, does not by itself guarantee that the operation is valid in the state the robot is actually in. Nor does it guarantee anyone will check whether it worked.

EmbodiedSkills reframes every skill decision as an execution proposal rather than a command. Before the skill runs, the runtime checks its prerequisites. After it runs, the runtime verifies the outcome. A decision is something to be tested against the world on both sides, not simply carried out. The two checks answer the two gaps in the problem directly. The prerequisite check asks whether this operation is valid in the current state. The verification asks whether the outcome actually happened. Neither question is left to the model that proposed the skill in the first place.

The piece that holds this together is a shared executable-skill interface. It connects three layers inside a single agent loop: high-level skill selection, bounded low-level execution by a VLA policy, and verification after the action. Because that interface stays fixed, the low-level VLA policies can be swapped out or adapted without touching the loop around them.

The interface does a second job as well. It records planning, execution, verification and recovery events as structured trajectories. Those records provide supervision for individual components, and when interactive feedback is available they can support optional online adaptation. So the loop that runs the robot is also the source of data for improving its parts.

The framework is instantiated with Qwen3-VL and with OpenPI, specifically pi0.5, and tested on RoboTwin 2.0 and LIBERO. Task-adapted low-level VLA policies reach an average success rate of eighty-six point two percent across fifty RoboTwin tasks. On the four LIBERO suites the average is ninety-seven point four percent.

The authors are careful about what these numbers establish: the execution performance of the task-adapted low-level policies used inside EmbodiedSkills. A harder test comes from four memory-dependent tasks in RMBench, where the same task-adapted execution approach achieves twelve point five percent average success. That gap between the two sets of results is part of the report, not hidden from it.

What is on offer is a layer rather than a policy. The VLA models underneath do the moving. EmbodiedSkills sits above them as an agent layer that is trainable and inspectable, checking before and verifying after, and recording everything it does. The stated aim is to turn capable low-level policies into closed-loop embodied systems, where each step is proposed, checked and confirmed instead of simply predicted.