A crucial piece missing from the foundation model for Embodied AI (EAI) is plasticity—the ability of continual learning without human intervention. While the emergence of In-Context Learning (ICL) has been pivotal to the success of Large Language Models (LLMs), its limitations and underlying mechanisms remain underexplored. This study illuminates the potential of large-scale meta-training, which prioritizes acquiring general-purpose ICL capabilities over mastering specific skills. We believe this technique could form a cornerstone of the next generation of general-purpose foundation models for EAI. Additionally, we introduce two open-source projects that are designed to advance the development of these foundation models.
