Two Days Watching 3D Picking on a Customer's Production Line — Some Practical Takeaways

The client does logistics sorting and has deployed AIOI SYSTEMS' general automated picking solution, with Eureka chosen for the vision part. I tagged along with their engineers on the production line for two days to watch this 3D vision picking system run.

First, what this thing does. 3D vision picking, in plain terms, is letting a robotic arm, from a messy pile of parts, recognize on its own which one to pick and where to grasp it, without pre-arranging the parts neatly or making dedicated fixtures. Eureka's setup is a pure vision solution; hardware-wise it's just a 3D camera plus algorithms, adaptable to different brands of robotic arms.

The first day was mostly watching. The engineer opened their software; on the left is the 3D point cloud image, on the right are the grasping strategy parameters. The point cloud refreshes in real time, and the outlines of stacked parts are quite clear. He demonstrated the calibration process—camera calibration plus hand-eye calibration—which takes about twenty minutes to go through. I thought this step was smoother than expected; with some vision solutions I've encountered before, calibration required repeated parameter tuning, but with Eureka you basically just follow the wizard.

The second day I tried changeover myself. The client's line was running glossy side mirror covers, which are highly reflective, and ordinary 3D cameras produce point clouds with missing chunks.

Eureka uses AI for completion, and after completion, grasp points can be generated normally. But I noticed a detail: the accuracy in the completed area is lower than in directly captured areas. The engineer said that for complex parts they adjust the grasping strategy to be more conservative—better slow than a failed grasp.

There are pitfalls too. Changing product models requires re-collecting point clouds and re-calibrating the grasping strategy; for an experienced engineer this takes about half a day, for a novice maybe a full day. The material says typical line cycle time is 4 to 8 seconds, picking success rate 98% to 99.9%. From what I saw, it's about right—simple parts with suction cups take about 5 seconds, complex parts with precise grippers are slower. But that's data after stable operation; the half-day of changeover is non-productive.

Another point: this system still has a baseline requirement for incoming material regularity. Completely disordered, mutually jammed stacks—it will also fail to pick them out, requiring a vibratory bowl or manual intervention. The marketing says no sorting needed, but in reality it's not needed in most scenarios, not all scenarios.

Whether it's worth deploying, my judgment is to look at batch size. For high-volume, low-variety lines with stable incoming shapes, this vision solution can indeed save labor, and the claim of recouping investment in a year isn't exaggerated under that premise. For low-volume, high-variety scenarios with daily changeovers, the calibration time amortized makes it uneconomical. Also, it's best to have an engineer on the team who understands vision, otherwise when problems arise you can only wait for original factory support.

After two days of watching, it can indeed make disordered picking stable in matching scenarios; in mismatched scenarios, it's useless.