EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents
EmbodiedSkills wraps robot skill choices in runtime checks, bounded execution, and outcome verification.
The paper frames VLA action prediction as insufficient for long-horizon robot tasks because the agent must confirm that a proposed operation is valid and that it worked. Its fixed executable-skill interface links high-level skill selection, low-level VLA execution, verification, and recovery in one loop. The authors instantiate it with Qwen3-VL and OpenPI/pi0.5, reporting 86.20% average success on 50 RoboTwin 2.0 tasks and 97.40% across LIBERO suites for the task-adapted low-level policies. On four memory-dependent RMBench tasks, the same execution approach reaches 12.5% average success. HF Daily Papers' note
The paper frames VLA action prediction as insufficient for long-horizon robot tasks because the agent must confirm that a proposed operation is valid and that it worked. Its fixed executable-skill interface links high-level skill selection, low-level VLA execution, verification, and recovery in one loop. The authors instantiate it with Qwen3-VL and OpenPI/pi0.5, reporting 86.20% average success on 50 RoboTwin 2.0 tasks and 97.40% across LIBERO suites for the task-adapted low-level policies. On four memory-dependent RMBench tasks, the same execution approach reaches 12.5% average success. HF Daily Papers' note
score 5