Efficient Continual Adaptation of Large Robotic Policy with Online Meta-Learned Adaptors

Recent research on large robotic policy models pretrained across diverse tasks has demonstrated that these models can adapt to new tasks with significantly fewer demonstrations as compared to learning from scratch. This reduced need for in-domain data makes these large models excellent backbones for continual adaptation—a crucial capability for general autonomous agents. For instance, a household robot must adapt to a range of unseen tasks as required by every household. Conventional methods for continual adaptation often require tuning the entire model's parameters, rendering them impractical for large robotic policy models. To address this, adaptors from the field of Large Language Models (LLMs) have been shown to effectively adjust large models using a relatively small number of parameters. These adaptors are task-specific, which helps avoid common issues such as catastrophic forgetting during continual adaptation. However, directly applying these adaptors does not facilitate knowledge transfer from previously learned tasks. To overcome this limitation, we propose Online Meta-Learned Adaptors (OMLA) for the continual adaptation of large robotic policy models. OMLA enables knowledge transfer through a novel meta-learning objective. We validated the outperformances of the proposed approach through extensive experiments in both simulated and real-world environments. These findings suggest that OMLA offers a promising direction for efficient and continual adaptation in large robotic policy models.