Jiacheng Xu

We introduce an RL framework that unify the training of answer generation and verification in a single model.

We introduce a novel object selection mechanism to allow sim-trained policies to rapidly adapt to real-world visual perturbations.