Event

ZeroDex Achieves Zero-Shot Long-Horizon Dexterous Manipulation via Multi-View 3D Grounding

1 reports · 1 sources

Jisoo Kim et al. propose the ZeroDex framework, which grounds language instructions from calibrated multi-view RGB images into executable 3D task plans; execution is driven by 3D grounding rather than learned policies, requiring no demonstration collection or post-training for specific tasks or embodiments. The paper claims that in real-world experiments, its 3D grounding accuracy and task success rate exceed those of zero-shot visual prompting and constraint-based baselines using a single RGB-D image, and that the demonstration-free system outperforms a VLA policy fine-tuned with 100 task demonstrations; in long-horizon manipulation, execution verification and replanning can recover from intermediate failures, and it generalizes to tool-use tasks involving unseen objects and new scenes.