Two Papers That Actually Get Why Library Robots Keep Crashing Into Chairs

I spent a good chunk of last Tuesday reading arXiv papers, which is what happens when you're semi-retired and it's raining. Two caught my eye, both dealing with problems I watched engineers struggle with for years: getting robots to navigate real indoor spaces without embarrassing themselves.

The first paper comes from a team working with Unitree's Go2 quadruped in an actual library. Not a lab. Not a simulation. A library with readers, chairs, bags, and all the chaos that implies. When I was at Kuka, we had a client who wanted mobile robots in their corporate library. Beautiful space, terrible for robots. Narrow aisles, students leaving backpacks everywhere, those rolling book carts that appear out of nowhere. We never got it working reliably. The project quietly died.

This arXiv paper reports 100% navigation success in static scenes, dropping to 88% when things get crowded. Those numbers sound modest until you've tried it yourself. The team used RTAB-Map for their SLAM, Nav2 for planning, and the usual sensor fusion stack. Nothing exotic. What I like is they're honest about why they chose a quadruped over wheels: floor transitions, temporary clutter, partially blocked passages. Real problems. Not the sexy stuff that gets funding, but the stuff that actually kills deployments.

Their map accuracy came in at 3.7 cm mean error against surveyed control points. That's workable. Not perfect, but workable.

The second paper tackles something I've been complaining about for years: cheap LiDARs that lie to you. Budget 2D LiDARs, the kind you find on educational platforms and cost-conscious industrial builds, don't have intensity channels. So when they hit glass or a mirror or someone's shiny laptop bag, they just... make something up. The robot has no idea it's getting garbage data.

The team behind SENTINEL built a system that gives these sensors a reliability score between 0 and 1. When the score drops, the robot knows to fall back on wheel odometry instead of trusting corrupted scans. I'll be honest, this is the kind of unsexy infrastructure work that makes me optimistic about the field. Nobody's going to put "robot learns when its sensors are lying" in a press release, but it matters.

They tested on a GEFIER R1 (had to look that one up, it's a four-wheel skid-steer platform) with an RPLidar A2M12 and a RealSense D435i. The test arena was small, roughly 185 by 245 cm, with controlled failure elements including glass, mirrors, and shiny paper. Not a massive validation, and they're upfront about that. But the spatial reliability maps showed clear separation between clean data and garbage. That's the hard part.

What connects these papers is a certain honesty about limitations. The library robot doesn't claim 100% success in crowded conditions. The SENTINEL team admits their validation is hardware-only because these failure modes don't exist in simulation. Neither paper is trying to sell me a revolution.