ShadowGrasp: Image-based 6-DOF Grasping by Capturing Single View under Variable Lighting with Shadow Ray Supervision

Robotic research is confronted with significant challenges in grasping objects with diverse shapes, materials, and surface properties. In particular, mutual occlusions between objects in cluttered scenes can lead to limited sensor perception, which in turn affects the success rate of grasping. To reveal occlusions under limited viewpoints, inspired by shadow rays completing object shapes, this paper introduces a novel approach called ShadowGrasp. This method depends on single-view shadow images under variable light source to perceive the 3D surroundings of cluttered objects. We utilize a ray supervision method similar to neural radiance field (NeRF) to learn the scene representation. Compared to other NeRF-based 6-DOF grasping methods, our method uses single-view observation instead of multi-view, reducing the computational overhead for the robot during the scene perception phase. Through simulation experiments, we validate the advantage of our method in handling object grasping tasks in cluttered scenes compared to baseline methods.