Youcheng Xu

Wrote the hand-eye calibration for a DUCO GCR5 arm with a fixed RealSense D435i, using a touch-probe method, then ran the lab's SAM3 + AnyGrasp pipeline end to end on the real arm.

• 1.8 mm calibration residual (8 points), 7.9 mm between two independent sessions
• Three successful closed-loop grasps over two days
• Fixed a wrist-flip failure by replacing linear moves with joint moves from URDF IK

Adapted the GELLO teleoperation arm to a 22 kg industrial arm. Matched GCR5 and UR5 joint axes from their URDFs so the UR5 kit could be reused, then did assembly, servo bring-up and wiring. Teleoperation of the GCR5 now runs.

• 7 Dynamixel XL330 servos on one bus, ~52 Hz read with zero dropped packets

Collected bimanual teleoperation demos, converted them to LeRobot format, and set up π0.5 post-training with openpi. Found the wrist camera was overexposed in most frames, measured it, and re-collected the dataset with the problem fixed.

• Wrist-camera overexposure 71–85% → 0% after re-collection

Open-source cat toys are open loop: none of them watch the cat. The hard part is a reward signal, detecting from video when a cat is really engaged. Plan: an overhead camera, a few discrete play primitives, behaviour cloning from my own demos, then on-robot learning. Five cats give a built-in test for unseen individuals.

Assembled an SO-101 leader/follower pair; teleoperation works. Next: record demos, train with LeRobot, measure success rate, then build an XLeRobot (two arms on a mobile base).