Figure 1.

The pipeline of the framework. The motion capture system reconstructs the hand–object-pose $T_{\mathrm{object}}^{\mathrm{hand}}$, and then the model predicts the final grasp pose $P\left(g_{\mathrm{object}}^{\mathrm{hand}}\mid T_{\mathrm{object}}^{\mathrm{hand}}\right)$ based on $T_{\mathrm{object}}^{\mathrm{hand}}$. The green hand in the lower left figure represents the correct prediction.