Real-Time Robotic Grasp Stability Detection via DynConv-MIL: A Lightweight Deep Learning Approach

With the advancement of tactile sensing technologies, robotic grasp stability detection has become essential for ensuring reliable manipulation and operational safety in industrial automation. Accurately and efficiently detecting grasp instability is critical to prevent failures in real-time robotic tasks. To address the challenges of modeling complex, non-stationary temporal dynamics and detecting subtle variations in tactile signals while maintaining high accuracy and low inference time, we propose a Dynamic Multi-scale Convolutional Multiple Instance Learning framework (DynConv-MIL) for robotic grasp stability detection using only tactile time series data, without relying on visual input. DynConv-MIL integrates deformable patch embedding, multi-scale depthwise convolutions with structural re-parameterization, and an attention-based MIL pooling mechanism. Our model enable the model to capture both local and global temporal dependencies while preserving computational efficiency. Extensive experiments on benchmark datasets show that DynConv-MIL outperforms baseline models by achieving the highest detection accuracy and F1 score along with the lowest inference time, highlighting its suitability for real-time deployment in practical robotic applications.