It Can Carry Tens of Kilograms, but a Collapsed Cardboard Box Stops It

This Sainade iLoabot-M — my opinion is, recommend it to warehouses with steady volume, regular SKUs, and a genuine inability to hire loading dock workers. If your SKUs change three times a day and the boxes are soft and have been rained on, don't deploy it yet — even if you do, a human will just be cleaning up after it.

I'm not from a warehousing background; I do grasping and tactile sensing, and what I usually stare at is those last few centimeters at the end effector. So looking at this machine, my attention was entirely on how it gets boxes out of the truck container — I didn't look closely at navigation and the chassis.

In logistics, the automation rate for sorting, warehousing, and packaging is not low — it's the loading/unloading segment that's most reliant on people. The industry figure is around 20% automation, with the remaining 80% all piled onto those last twenty meters. In summer the inside of a truck container can hit 60°C; cold storage is -18°C. Both scenarios sound miserable to me, let alone spending an eight-hour shift inside one.

iLoabot-M was shown at the 2023 CIIF, and this year at WRC it's version 2.0. Its public claim is that it doesn't require modifying trucks or docks — it can handle containers, box trucks, and flatbeds, and both boxed and bagged goods. I believe half of that. Not modifying the dock is true, but "works on any truck" depends on whether the container's inner walls are clean and the floor is flat; if an old truck is warped, it'll be just as lost.

The on-site engineer handed me the host computer; the interface isn't too complex. On the left is the container point cloud, on the right is the task queue, and in the middle you select the truck type, frame the work area, and set stacking rules. I went through it as he described: scan the container, frame the range, dispatch the task.

The first snag was the point cloud. The container interior is dim, and the cardboard colors are similar; the seam between two boxes pressed together was almost invisible on screen. It couldn't segment them on its own, so in the end I manually drew a couple of lines to split the area. The engineer said this is standard practice — I get it, but for a frontline operator, this is no longer "one-click start."

The grasp itself was steadier than I expected. The end effector swaps different heads for boxed vs. bagged goods; the one I saw was equipped with a vacuum type — press on, evacuate air, lift, very crisp motion. The first box wobbled a bit when it came up, center of gravity off; it set it down and re-grasped once and it was straight. I'll give it credit for this: it didn't force the lift, it probed first then proceeded — the logic is right.

There were three types of goods that were tricky on site.

Crushed boxes. Cardboard that's been damp or crushed has a dent on top, and the suction cup leaks when pressed on. Vision can see the dent, but it doesn't know whether that area is big enough to grip. In that case you either switch to a clamping method or have a human go steady it.

Edge-hugging. The last row of boxes is pressed right against the container wall, no room to reach in, so you have to go in at an angle. This motion places high demands on arm configuration; on site I watched it try twice before hooking the box corner out.

Bags. Soft packages have no fixed shape, and the grasp point is different every time. After swapping the end effector it can grasp them, but the cycle time clearly slows down.

The weak spot in this kind of scenario is contact sensing at the moment of attachment. Vision gives pose; force and tactile sensing are what tell you whether it's actually stuck, whether it's leaking, whether the box is slipping in the gripper. Right now most robot makers still rely on vacuum switches plus flow-rate judgment — good enough, but not fine enough. I just started getting into tactile arrays these past two days, and that's exactly the direction I'm going: thin-film sensors attached to the suction cup skirt, which in theory could turn attachment area into a readable number instead of only finding out after a box drops. But that's a lab-stage idea, not yet verified on a real machine.

One more thing: the cost of switching SKUs. I asked on site how long it takes to switch product specs, and the answer was re-collecting data and tuning parameters. That time can't be saved — everyone's solution is the same, it's just a matter of how long.

For warehouses with steady volume, few product specs, and genuinely hard hiring — especially cold-chain and high-temperature container scenarios that people don't want to enter — it's worth it. For those with SKUs changing daily, cramped sites, and boxes from all sorts of sources, don't rush; first run one line for three months and see the proportion of human backup.

Going forward, what I most want to see is how long it takes to switch product specs once — that number is worth more than the grasp success rate.