Vision-Language-Action (VLA) models have become a powerful approach for robot manipulation, but training a single generalist policy that works across heterogeneous robot embodiments remains a significant challenge. Existing methods often underutilize shared dynamics priors from diverse visual and interaction data, limiting cross-embodiment transfer. They also require extensive manual preprocessing to align different action spaces into a common format.
To address these issues, researchers introduce DyPES-VLA, a cross-embodiment VLA that learns shared Dynamics Priors and Embodiment-Specific control. The model first trains a vision-language model (VLM) with a future-prediction objective on cross-embodiment data, enabling the shared query representation to capture object motion, contact, and interaction-induced scene changes. This allows the model to learn dynamics priors that are transferable across different robots.
Second, an embodiment-specific Mixture-of-Experts (MoE) action head translates these shared dynamics priors into executable controls directly in each embodiment's native action space. This head shares attention layers to capture common temporal action structures, while its embodiment-specific feed-forward experts handle unique kinematic constraints and control semantics. This design eliminates the need for manual pre-alignment of heterogeneous actions.
As a generalist policy, DyPES-VLA achieves state-of-the-art performance across simulation and real-world evaluations, reaching 98.0% success on LIBERO, 59.25% on RoboCasa-GR1, and 89.02% on RoboTwin 2.0. The results demonstrate the effectiveness of learning shared dynamics priors and embodiment-specific control for cross-embodiment manipulation.