Wen-Qian Li

Frontier VLMs excel at precise scene estimation but falter in executing coordinated, goal-directed actions, highlighting a critical bottleneck for robot generalism.

VLMs exhibit distinct failure modes under physical visual stress, revealing that traditional accuracy metrics can mask critical vulnerabilities in embodied AI systems.

By intelligently reorienting adversarial style perturbations, SRasP achieves state-of-the-art cross-domain few-shot learning performance, demonstrating the power of stabilized adversarial training for domain generalization.