Preprint
Aug 2026
CometVLA: Co-Training on an Embodied Data Pyramid towards Physical Understanding
Results demonstrate that physical understanding pre-training genuinely benefits downstream manipulation inVision-language-action models, and introduce GAP tokens, a compact learnable bottleneck that isolates task-agnostic motion regularities and lets the action head consume physical commonsense without corrupting the p...
Hanwen Wan, Dafeng Chi, Lin-Bo Zhai et al.
· 1 citation