Georgia Gkioxari

Multimodal MoE experts spontaneously develop semantic specializations, enabling data-free identification and highly efficient, targeted adaptation that outperforms LoRA.

Camera-controlled world models don't need weight updates to master scene physics: optimizing a handful of test-time tokens via cross-attention outperforms LoRA and block fine-tuning at decoupling local object dynamics from global camera trajectories.

HANDOFF redefines humanoid robot control by seamlessly integrating diverse manipulation skills into a single, intuitive command interface.