Generating Robot Trajectories with Language Models: A Survey of Recent Approaches

In this survey paper, researchers explore the potential of utilizing large-scale pre-trained models (foundation models) in Artificial Intelligence to create general-purpose robots that can adapt to any environment and perform diverse tasks. These models have shown impressive open-set performance and content generation capabilities in fields like Natural Language Processing and Computer Vision. The authors survey various approaches to applying these models to robotics, including a model-agnostic (model-free) approach and direct employment of Large Language Models as dynamic models. However, they acknowledge the limitations of these approaches and highlight the need for more research on how humans and animals learn, which is efficient in terms of data and energy. The authors also emphasize the importance of grounding in multiple sensory modalities and embodiment perspective to achieve truly adaptable robots.