Cortex: A Bidirectionally Aligned Embodied Agent Framework for Long-Horizon Robotic Manipulation

Summary

Cortex is a bidirectionally aligned embodied agent framework designed to enable vision-language-action (VLA) models to handle long-horizon robotic tasks. While VLA models show promise as general manipulation policies, their Markovian nature—relying only on the current observation—limits performance on extended multi-step tasks. Hierarchical dual-system approaches address this but introduce a gap between high-level planning semantics and low-level execution kinematics. Cortex bridges this gap with a customized planning interface that communicates executable and tractable subtask plans from a high-level VLM to a low-level VLA. The framework standardizes manipulation subtasks into 32 canonical skill primitives and injects tractability principles, such as representative object attributes and improved trajectory reachability, into the data generation pipeline. Experiments show improvements of 3.1% over monolithic baselines on Libero-long and 4.1% on RoboTwin. Notably, Cortex's general VLM can zero-shot complete unseen real-world long-horizon tasks, such as multi-stage chemistry experiments, by simply combining fine-tuned VLAs. The paper is available on arXiv as 2607.05377.