Lifelong Morphology Learning for Deformable Embodied Agents

Abstract

A deformable agent can continuously adjust its morphology during training, allowing it to discover more suitable structures and outperform fixed-morphology counterparts in terrain-specific tasks. This adaptability is achieved through a joint optimization process consisting of two stages: the Skeleton Transform stage which modifies the agent’s morphology and the Execution stage which optimizes the control policy. However, enabling a deformable agent to continuously learn new policies for different terrains without forgetting previous tasks remains a major challenge. Continuous terrain changes can easily disrupt previously learned strategies, making it difficult to adapt to new tasks while maintaining performance on earlier ones. In this work, we focus on lifelong morphology learning for deformable agents that must adaptively traverse a sequence of diverse terrains. We propose Ske-Ex, a lifelong learning framework where both the Skeleton Transform and Execution stages are designed for lifelong adaptation. Unlike existing methods that optimize only control policies under fixed morphologies, Ske-Ex supports joint adaptation of structure and control, making it better suited for deformable agents. We adopt a regularization-based approach as our lifelong learning strategy, as it avoids the need to store large amounts of prior task data. Experimental results show that Ske-Ex exhibits strong resistance to forgetting and superior generalization, and that the joint optimization of both modules outperforms using either stage alone. Additionally, we introduce a flexible MuJoCo terrain benchmark to facilitate future research on lifelong learning for deformable agents. Our demonstration videos are available at https://johncenavsbatista.github.io/Ske-Ex/