Improving Multimodal Interactive Agents with Reinforcement Learning from Human Feedback