Knowledge Diversion for Efficient Morphology Control and Policy Transfer

Universal morphology control aims to learn a universal policy that generalizes across heterogeneous robot morphologies, with Transformer-based controllers emerging as a dominant choice. However, such architectures incur substantial computational costs, resulting in high deployment overhead, and existing methods exhibit limited cross-task generalization, necessitating training from scratch for each new task. To this end, we propose DivMorph, a modular training paradigm that leverages knowledge diversion to learn *decomposable controllers*. DivMorph factorizes randomly initialized Transformer weights into *basic knowledge units* via SVD and employs dynamic soft gating, conditioned on task and morphology embeddings, to adaptively modulate these units into universal *learngenes* and morphology- and task-specific *tailors* during training, thereby achieving knowledge disentanglement. By selectively activating relevant components, DivMorph adaptively recomposes the controller, enabling efficient policy deployment and effective policy transfer to novel tasks. Extensive experiments demonstrate that DivMorph achieves state-of-the-art performance, improving sample efficiency for cross-task transfer by 3.3$\times$ and reducing model size for single-agent deployment by 16.7$\times$.

Modern robots come in many different forms, ranging from simple walking robots to complex systems with many limbs. Building a single control system that can operate across these diverse robot bodies and also adapt to new tasks remains a major challenge in robotics. Current methods often rely on large neural networks that are costly to deploy and usually need to be retrained from scratch whenever the robot encounters a new task. To address this problem, we propose DivMorph, a training framework designed to learn reusable and modular control knowledge. Rather than learning one large controller for all situations, DivMorph breaks the controller into smaller functional components. These components can then be flexibly combined according to the robot’s body structure and the task it needs to perform. During training, the framework automatically learns which knowledge should be shared across different robots and which parts should remain specialized for particular morphologies or tasks. This modular design enables robots to reuse previously learned skills more efficiently, making adaptation to new tasks faster and reducing the computational cost of deployment. As a result, DivMorph can generalize more effectively to unseen tasks while requiring much smaller models for individual robots. Experiments on a wide range of robotic control benchmarks demonstrate that our method substantially improves transfer efficiency and significantly reduces model size compared with existing approaches.