Abstract
Accurate grasp stability assessment is essential for reliable robotic manipulation. However, most approaches treat grasping as a static action, overlooking dynamic challenges from multistage interactions, sensor noise, and the lack of physical commonsense. To address these limitations, we propose multistage dynamic multi-modal grasp (MDM-Grasp), a framework for grasp stability assessment. The framework employs a cascaded two-phase architecture that integrates spatiotemporal tactile sequences, visual inputs, and prior physical knowledge extracted from a large language model (LLM). To enhance robustness under realistic conditions, we introduce a mutual cross-attention fusion mechanism (MCFM) for effective integration of heterogeneous and noisy sensory streams. We further construct a comprehensive dataset encompassing diverse lighting, object, and sensor degradation conditions. Extensive experiments demonstrate that MDM-Grasp achieves state-of-the-art performance, with F1-scores of 99.19% in the initial assessment phase and 96.57% in the refinement phase and generalizes well to unseen objects with a 96.67% real-world success rate. These results highlight the potential of combining multimodal sensing with language-model-based priors to enable more reliable and human-like robotic grasping in complex environments.