Neuromeka and KAIST drop the wrist camera as one head-mounted RGB camera guides humanoid EIR through two-handed cup separation

Korean cobot maker Neuromeka said on Oct 2 that it and the research group of Professor Kim Min-jun at KAIST's School of Electrical Engineering developed SCDP (Spatially Conditioned Diffusion Policy), a control method that lets its EIR humanoid do fine two-handed work using only a single RGB camera in its head, Robot Newspaper (irobotnews) reported; the paper was accepted at the Conference on Robot Learning (CoRL) 2026, and MEKAWIRE has not yet seen the paper or a company release, so the figures below are as reported.

SCDP predicts where the robot's fingertips will travel, projects that path onto the camera image and pulls visual features from several resolutions around the expected contact area, re-selecting them as the motion is refined.

According to the report, EIR separated a stack of nested paper cups with both hands without any wrist camera, and in Meta-World simulation with 20 demonstrations per task SCDP averaged 82.5% success on the six hard tasks against 74.2% for a standard diffusion policy that also had a wrist camera; in real cup-grasping trials on a Franka Panda arm with visual clutter it scored 85% (handle) and 80% (wall) against 30% and 65% for the baseline, over 20 trials each.

Neuromeka CTO Heo Young-jin said removing wrist cameras and cabling frees space around the hand for work with human tools in tight cells. The research follows Neuromeka's role as lead organisation in a KRW 17.5bn science-ministry project on autonomous shipyard welding, which MEKAWIRE reported on Sept 28, and comes five days before Tomorrow Robotics published its Habilis Brain 0 manipulation model.