Large Language Models as Generalizable Policies for Embodied Tasks

Large language models can be adapted through reinforcement learning to serve as generalizable policies for embodied visual tasks, outperforming other common baselines in terms of paraphrastic robustness and behavior generalization.