Paper Overview
Research area: Computer Vision / Embodied AI Authors: Jiaqi Peng, Xiqian Yu, Delin Feng, Yuqiang Yang, Wenzhe Cai, Jing Xiong, Ganlin Yang, Jinliang Zheng, Jiafei Cao, Xueyuan Wei, Jiangmiao Pang, Yuan Shen, Tai Wang Published: 2026-07-06 arXiv: 2607.05377
Summary
Recent vision-language-action (VLA) models show promise as general-purpose manipulation policies, but they struggle with long-horizon tasks because of their Markovian nature — relying solely on the current observation. Hierarchical dual-system approaches address this issue, yet a gap remains between the semantics of high-level planning and the kinematics of low-level execution.
This paper proposes Cortex, a bidirectionally aligned embodied agent framework featuring a tailored planning interface that communicates executable and tractable subtask plans from a high-level VLM to a low-level VLA. Key design elements include:
- Standardizing manipulation subtasks into 32 canonical skill primitives
- Injecting tractability principles (e.g., representative object properties and improved trajectory reachability) into the data generation pipeline
- Libero-long: +3.1% over monolithic baselines
- RoboTwin: +4.1% over monolithic baselines
- Cortex's general-purpose VLM achieves zero-shot completion of unseen real-world long-horizon tasks (e.g., multi-stage chemistry experiments) by simply combining it with a fine-tuned VLA.
Results
*Auto-collected on 2026-07-06.*