Mink-GraSNet: computationally efficient 6-DoF grasp detection for large-scale AI-powered robotics

Mink-GraSNet: computationally efficient 6-DoF grasp detection for large-scale AI-powered robotics

연구 분야: Artificial Intelligence

학회: The Journal of Supercomputing

Achieving precise and robust 6-DoF robotic grasping in complex, large-scale cluttered environments presents a significant computational challenge for AI-powered robotics. Processing high-resolution 3D point clouds for real-time grasp detection is often computationally demanding, hindered by methods that struggle with large data volumes and require high latency. To address these limitations and enable efficient, scalable 6-DoF grasp detection, we propose Mink-GraSNet, a novel end-to-end deep learning framework. Mink-GraSNet leverages graspness as a core guiding principle, combining sparse geometric-semantic feature fusion with an adapted efficient sparse convolutional backbone, MinkUNeXt. Crucially, we introduce SACA (Sparse Adaptive Channel Attention), an efficient module specifically designed for sparse tensors to enhance feature discriminability without sacrificing efficiency. This collaborative approach achieves a strong balance between powerful feature representation and high computational efficiency. Experimental results on the large-scale GraspNet-1Billion benchmark demonstrate state-of-the-art grasp detection accuracy. Furthermore, our method exhibits significantly faster inference speeds, showcasing its capability to efficiently handle substantial computational demands and suitability for high-performance robotic applications.