ZeroDex: Zero-Shot Long-Horizon Dexterous Manipulation via Multi-View 3D-Grounded VLM Reasoning
ZeroDex uses a vision-language model to turn language instructions into executable 3D task plans from multi-view images. It grounds instructions by finding reference frames and keypoints, then fuses these views into consistent 3D data. This enables zero-shot manipulation tasks like pick-and-place and tool use by using modular reasoning and reusable action primitives.