Robots in real-world environments face three challenges: requiring many expert demonstrations, lacking clear task guidance, and facing safety risks from failures. We propose three solutions: SCAN, which selects demonstrations aligned with task progress; VICtoR, which uses contrastive learning and vision-language models for dense guidance; and PrObe, which detects anomalies via feature consistency. These methods were accepted in top AI conferences AAAI, NeurIPS, and ICLR, and won multiple awards.

This work addresses three key challenges in real-world robotics and was evaluated against the state-of-the-art: (1) a few-shot imitation learning model with attention boosts performance by 27%; (2) a hierarchical task guidance model combining language and vision improves task efficiency by up to 43%; (3) a new anomaly detection task, benchmark and contrastive detection method achieve top AUROC in 17 out of 21 tasks.

This technology covers few-shot imitation, (LLM-augmented) task guidance generation, and anomaly detection—key enablers for accelerating real-world robot deployment with strong industrial potential. As commercial robots face diverse, complex scenarios, the first two modules enable fast adaptation and decision-making. With real-time anomaly detection, the system enhances safety and robustness. These complementary innovations collectively drive the deployment of intelligent robotic systems.