A week watching PUDU D9 on site: has it cleared the bipedal hurdle?

A friend recommended PUDU's PUDU D9 to me, saying a warehouse distribution client of theirs in South China was trialing it, so I asked around and arranged a visit. Let me be clear upfront: this thing is 170 cm tall and 65 kg, not something I can haul home to test, so this week was three encounters — a booth demo, half a day squatting at the client site, plus a sales-provided operation video and spec sheet to cross-reference.

The first day was mainly about whether it can walk. First impression is the height, 170 cm, about the same as me. The engineer said it has 42 degrees of freedom total, which simply means 42 movable joints, with max joint torque of 352Nm. I did the math — pushing a fully loaded tote shouldn't be a problem.

The demo covered upright walking, someone pushing it from the side, going up a small slope, and finally walking over to operate PUDU's own SH1 floor scrubber for floor cleaning. It walks steadily, but what I watched was those two seconds after it stops — the feet do little shuffling steps to find balance. Has it actually run on a production line? That's usually my first question. Demos don't count.

The third day I went to the client site to see where it gets stuck doing work. The engineer used a tablet for task orchestration, dragged a few waypoints, and had it walk to a shelf, grab a cardboard box, and hand it over. The hand is that 11-DOF five-finger dexterous hand, the PUDU DH11. Grabbing rigid cardboard boxes is fine, but when grabbing soft packaging the fingers close a bit hesitantly.

Navigation relies on visual semantic navigation, meaning the camera understands "this is a shelf" and "that's an aisle entrance," then decides where to go. This thing is sensitive to environmental changes. There was a reflective patch on the floor, and it stopped there twice. Later a shelf was moved half a meter, and the waypoints had to be re-marked. I've been using vision cameras for three weeks, so I know this all too well — re-labeling data when switching product models saves you zero effort.

After a week, here's my conclusion. The good: bipedal stability and torque are sufficient, it can climb slopes and resist pushes, the dexterous hand's grasp range is wider than a gripper, and it can coordinate with its own SH1 and similar equipment — it's not fighting alone. The bad is also obvious: changing scenes requires re-labeling, reflections and dynamic obstacles still need human supervision. I squatted there for half a day, and out of about ten grasps, roughly two needed the engineer to step in and steady it. The intervention rate is still on the high side.

The conclusion is: it depends. Suitable for fixed routes and low-SKU scenarios, like inspection, delivery, and coordinating with cleaning equipment; not suitable for throwing straight into mixed-line sorting. Right now it's more like a mobile platform that can stand up, not yet a workstation that can carry the load on its own.