SE(3) Neural Potential Fields for 6-DoF Trajectory Planning Directly from Images Without Explicit 3D Reconstruction
The paper reports image-trained robot motion fields that reached blocked grasp poses without collisions in two tabletop tests.
The method learns an SE(3) potential field from posed RGB images, using a navigation-function target during training instead of an explicit 3D reconstruction at planning time. In UR10 runs from obstacle-blocked starts, it converged within 3 cm of the grasp from every start and kept executed paths collision-free against ground-truth geometry. The reported grasp success was 90.0% in one scene and 40.0% in the other, with remaining failures attributed to the Cartesian executor. Planning took about 2 seconds, versus 67-133 seconds for RRT* on reconstructed scenes. ArXiv · AI/CL/LG's note
The method learns an SE(3) potential field from posed RGB images, using a navigation-function target during training instead of an explicit 3D reconstruction at planning time. In UR10 runs from obstacle-blocked starts, it converged within 3 cm of the grasp from every start and kept executed paths collision-free against ground-truth geometry. The reported grasp success was 90.0% in one scene and 40.0% in the other, with remaining failures attributed to the Cartesian executor. Planning took about 2 seconds, versus 67-133 seconds for RRT* on reconstructed scenes. ArXiv · AI/CL/LG's note
score 5