LiDAR Odometry Is Getting Seriously Good, and Here's Why That Matters for Warehouse Robots

When I was at Kuka, we spent a lot of time wrestling with exactly this class of problem on mobile platform projects. The IMU data was always dirtier than the spec sheets suggested, especially in industrial environments with vibration from nearby machinery. You'd correct for motion, and then you'd find your correction had introduced its own small errors. It was, to put it mildly, annoying. So I have some appreciation for what the AC-LIO team has pulled off here.

The second paper, FAST-LIVGO, goes a step further by fusing LiDAR, IMU, cameras, and GNSS together into one tightly coupled system. The interesting bit is how it handles situations where one or more of those inputs degrades or fails. In a tunnel, you lose GNSS. In a textureless white corridor (and anyone who's deployed robots in cold storage knows exactly what I'm talking about), your visual odometry can fall apart. FAST-LIVGO uses what they call a degeneracy-aware dual-mode outlier rejection strategy, basically a system that monitors how well the LiDAR-visual side is performing and switches to GNSS-aided recovery when it starts to struggle. They tested it on a custom dataset using a fixed-wing UAV flying at around 20 metres per second, which is a demanding scenario. Whether this translates cleanly to slower ground vehicles in industrial settings remains unclear, but the underlying architecture is sound.

The third piece of the puzzle is Odyssey, a new dataset from the University of Göttingen designed specifically for evaluating odometry systems in GNSS-denied environments. This one's less a novel algorithm and more a contribution to the research infrastructure. The headline feature is the ground truth system: a navigation-grade Ring Laser Gyroscope combined with RTK/INS. The bias stability is, according to the paper, one to four orders of magnitude better than what existing automotive datasets use. They've included 36 sequences across varied environments, and crucially, they've included prolonged indoor scenarios like parking garages where GNSS is simply gone for extended periods. That's been a real gap in publicly available benchmarks. Most existing datasets only deal with sporadic GNSS dropouts, not the sustained absence you'd encounter in a large underground or enclosed facility.

Now, look, here's the thing. These are academic papers. The gap between a promising arXiv result and a system that actually works reliably on a warehouse floor, day after day, in the presence of forklifts and dust and people who've taped cardboard over half the reflective markers, is substantial. I've seen enough promising lab demos turn into expensive disappointments to be cautious. It's too early to say which of these approaches will find its way into commercial products.

But the trajectory of this research, taken together, is encouraging. The localization problem for large-scale autonomous vehicles isn't solved, but it's getting less unsolved, which is sort of the best you can say about any hard engineering problem in progress. If you're specifying autonomous mobile robots for a new facility in the next year or two, the navigation accuracy question is worth asking your vendor about in some detail. The answers are getting better.

Robert "Bob" Macintosh · Yesterday · 4 min