Siyuan Huang

A massive 2.6M-scene benchmark and a novel generative model finally unlock robust dexterous grasping in cluttered environments, outperforming prior methods without post-optimization.

Robots can learn contact-rich dexterity directly from human skin: co-training on aligned human visual-tactile-action data more than doubles real-world manipulation success (70% vs. 29%) with zero performance saturation up to 1,000 demonstrations.

Adversarial clothing can successfully evade thermal person detectors with over 80% effectiveness, exposing critical security flaws in AI-driven surveillance systems.

DreamHand achieves a groundbreaking 40% reduction in error for 3D hand trajectory recovery in occluded environments, setting a new standard for egocentric video analysis.