Two Days With the RBQ Quadruped Robot at a Shipyard: Some Honest Thoughts

The trouble I've run into lately is: after the robot finishes a lap, where does the data go?

Last week I went to a client site — a heavy equipment yard — where they're trialing RBQ for inspection. I stayed two days, nominally to help check interfaces, but really I just went over this dog from head to tail. Let me be clear up front: I can't take this thing home. Everything below is from following along on site, watching engineers tune parameters, and then digging through the manual myself.

The RBQ series made this year's Korea Top 10 Mechanical Technologies list. The official line is that it can move stably over rugged and unstructured terrain, and can flexibly switch between dynamic gaits and reinforcement-learning gaits.

Prep is more fiddly than you'd think. First you have to map the site, draw the inspection route, and fix the auto-docking station's position. The body carries a 3D lidar, a depth camera, and a gimbal camera. Docking is automatic — sounds worry-free. But the mapping step is very site-dependent. In a yard, the stockpiles change daily. The map is dead, the materials are alive.

To get started you first teach it the route. The process isn't hard — like many inspection robots, you teach it once, save it as a task, and it runs on its own. The first day went smoothly; I even thought there was nothing worth writing about.

Then I got stuck on the manual.

They made the manual into a page you can ask questions. You ask it about interface parameters and it answers with total confidence. I asked twice about the trigger conditions for gait switching, and the two answers weren't quite the same, and didn't match the PDF version. For an engineer this is fatal — you can't tell which one is real. In the end I just read the static PDF and used that page as a search tool.

The second pitfall is gait switching. Dynamic gait and reinforcement-learning gait — officially, you can switch flexibly. From my testing, the dynamic one is more stable on wet slopes, while on gravel piles the learned gait actually hugs the ground better. But the switch isn't its own judgment — most of the time you have to choose. Choose wrong and the dog walks like it's drunk. Looks scary, but it's not actually broken.

One more thing: they say it can climb slopes blind, without vision. That capability is real — the engineer demoed it on the spot. But the problem is, when it's not using vision, the logs don't show which mode it was running in either. Troubleshooting relies entirely on human memory. Classic case of observability not keeping up.

There were pleasant surprises too. Docking succeeded on the first try, no getting stuck at the station. It ran two or three hours straight with no drift in posture. The hardware is solid.

My conclusion is a it-depends. Fixed outdoor-site security patrols, repetitive inspections in shipyards and heavy industry, and universities/labs wanting a platform for secondary development — those scenarios fit. Conversely, if you expect it to understand your site out of the box, if you have no manpower to maintain maps and routes, or if it's an indoor environment cluttered with junk — none of those fit.

As for whether it's worth adopting: first check whether you have someone on your side who can catch the data it spits out.