HTC-Grasp: A Hybrid Transformer-CNN Architecture for Robotic Grasp Detection

Accurately detecting suitable grasp areas for unknown objects through visual information remains a challenging task. Drawing inspiration from the success of Vision Transformer in vision detection, hybrid Transformer-CNN architecture robotic known as HTC-Grasp, is developed to improve accuracy grasping objects. The employs an external attention-based hierarchical encoder effectively capture global context and correlation features across entire dataset. Furthermore, channel-wise CNN decoder presented adaptively adjust weight channels approach, resulting more efficient feature aggregation. proposed method validated on Cornell Jacquard dataset, achieving image-wise detection 98.3% 95.8% each respectively. Additionally, object-wise 96.9% 92.4% same datasets are achieved based this method. A physical experiment also performed using Elite 6Dof robot, with rate 93.3%, demonstrating method’s ability real scenarios. results study indicate that outperforms other state-of-the-art methods.