libero-long-14d79fb2·1 events·first seen Aliases: Libero-long
Cortex is a new embodied agent framework that bridges the semantic gap between high-level Vision-Language Model (VLM) planning and low-level Vision-Language-Action (VLA) execution for long-horizon robotic manipulation. The system standardizes manipulation into 32 canonical skill primitives and uses an event-balanced sampling strategy to handle subtask transition ambiguity, enabling automatic annotation of over 4,000 hours of video data. Cortex outperforms monolithic baselines by 3.1% on Libero-long and 4.1% on RoboTwin benchmarks, and demonstrates zero-shot generalization to unseen real-world tasks such as multi-stage chemistry experiments.