Khoa Vo

Gripper-aware learning can drastically enhance robotic manipulation, with GVLA outperforming existing models in adaptability and performance across diverse tasks.

Advanced Vision-Language-Action models can be dramatically compressed by up to 50% without losing performance, reshaping our approach to robotic manipulation.

Separating appearance from identity in video object learning leads to significant gains in segmentation and temporal consistency, outperforming existing methods.

Rare but critical categories like children and emergency vehicles can be effectively detected using a novel semantic-guided expert framework that transforms long-tail learning in 3D perception.