Improving Robotic Grasp Detection Under Sparse Annotations Via Grasp Transformer With Pixel-Wise Contrastive Learning

Despite the tremendous success in robotic grasp detection, learning a robust grasping policy with limited annotations still remains significantly challenging, especially in complex industrial scenarios. It is nontrivial to effectively collect a large number of high-quality grasping annotations under real-world industrial scenarios. In this article, we first investigate the main causes of the performance collapse in prior methods under the sparse annotation setting, attributing it to multigranularity perception impairment and texture-centric feature overemphasis. Based on this analysis, we further propose an efficient sparse grasp detection framework, grasp transformer with pixel-wise contrastive learning (GT-PCL), which incorporates a shunted attention mechanism to explicitly synthesize multiscale features and extract geometric invariant representations as task-relevant grasping priors by restricting cross-view pixel consistency. Experimental results demonstrate that GT-PCL significantly outperforms current state-of-the-art methods with a considerable margin on three standard benchmarks and boasts an average 12% success rate gain in complex scenarios with a physical robot arm UR5.