Hongliang Li

Robots can now assemble novel products alongside humans, guided by natural language, thanks to a new LLM-powered framework.

VLMs forget less when learning new tasks if you explicitly preserve the semantic relationships they learned during pre-training, especially in vulnerable regions identified with adversarial anchors.

Adapt your action anticipation model on-the-fly to new viewpoints (egocentric or exocentric) with a novel test-time adaptation method that leverages multi-label prototype growing and dual-clue consistency.