Preprint drops camera-to-robot calibration from self-supervised industrial robot arm grasping

A new arXiv preprint proposes a self-supervised grasping pipeline that drops the camera-to-robot extrinsic calibration step from the vision-to-control loop. The work, accepted at ICCAS 2026, runs on a single UR5e with a gripper and a monocular wrist-mounted camera.

The mechanism: rather than hand-calibrating camera-to-robot pose, the robot auto-generates demonstrations around a target pose, and a convolutional network regresses target-relative translation and rotation directly from a single RGB frame. A coarse-to-fine controller applies a height-variation coarse model first, then a planar refinement model for final alignment.

In simulation, refinement cut final planar dispersion from 9.69 mm to 5.38 mm. On the real UR5e, end-to-end grasp hit 66.6% and 63.6% on two objects without rotation, with partial robustness under rotated conditions, across three physical objects total.

Everything is author-reported on a single preprint, one robot, three objects, and only the authors' abstract and section excerpts are available. The authors flag remaining challenges in depth prediction and object-dependent generalization, and no independent reproduction appears in the source bundle.

What remains unknown: whether the refinement gain survives cross-lab reproduction, broader object sets, or a different gripper, and whether the calibration-free claim holds on hardware the team did not test.