"EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents" by Wei Wang , Wenqiao Zhang , Yutong Lin , Yuqian Yuan , Tianwei Lin , Jinhao Mao , Zhenxuan Fan , Mingjian Gao , Yang Dai , Wentong Li , Zheqi Lv , Zheng Dong , Yingjie Niu , Jiaqi Zhu , Jun Xiao , Chao Li , Yueting Zhuang
TLDR:
EmbodiedSkills introduces a unified framework that validates and verifies robot skill executions using a fixed interface, enabling adaptive low-level vision-language-action policies within closed-loop embodied agents. While existing models directly link visual and language inputs with robot actions, the framework considers the need for perception, planning, execution, progress verification, and recovery in long-horizon tasks. By treating each skill decision as an execution proposal and including runtime checks for prerequisites and outcome verification, EmbodiedSkills ensures the validity of operations post-execution. Through a shared interface connecting high-level skill selection and low-level execution, the framework allows for adaptable policies without altering the agent loop. The method records events for structured trajectories, supporting individual component supervision and online adaptation with interactive feedback. EmbodiedSkills, applied with Qwen3-VL and OpenPI/pi0.5 on RoboTwin 2.0 and LIBERO, demonstrates successful task adaptation with high average success rates, establishing efficient execution of low-level VLA policies.
Read Paper / Blog
huggingface.co
Paper page - EmbodiedSkills: A Unified Framework for Orchestrating, Training, and Deploying VLA Agents
Join the discussion on this paper page
1September 8, 2026 48
