So I saw someone released a localization stack that supposedly figures out where the robot is from the very first scan, no pose hint needed. Sounded perfect for my setup because right now I have to drop my bot near a known spot every morning or AMCL just wanders around confused for a minute. I cloned the repo, built it, hooked it into my nav stack, and it runs fine but here's the thing. When I start the robot at the far end of my workshop it still publishes a garbage pose til I give it an estimate manually. Am I missing some config flag? The readme is pretty thin on the launch parameters side. For context my map was built a while back and the workshop has changed a lot since, moved shelving around, added a workbench, that kind of thing. Could a stale map be the reason it can't snap to where it actually is? Or does this thing expect a really clean enviroment with distinctive features everywhere? My lidar is a cheap 2D unit if that matters. woud love to hear if anyone else got this working hands free, because right now I can't tell if I misconfigured something or if my space is just too boring for it to lock on.
Why does my robot still ask for an initial pose after switching to that new localizer?
Nov 02, 2026
Last reply: 19 hours ago
6 Replies
First reply
Stale map is my first suspect, and honestly it's probably enough on its own to explain what you're seeing. These global localizers work by matching what the lidar sees right now against the map you built before. If you've moved shelving around and dropped a workbench in the…
Read the full replyStale map is my first suspect, and honestly it's probably enough on its own to explain what you're seeing. These global localizers work by matching what the lidar sees right now against the map you built before. If you've moved shelving around and dropped a workbench in the middle of the space, big chunks of every scan describe geometry that isn't on the map. The matcher has nothing solid to grab onto, so it either sits on a junk pose or waits for you to tell it roughly where to look. The cheap 2D unit makes it worse, not because it's bad gear but because a short-range spinning lidar sees a thin slice of the room. Long bare walls, rows of similar shelving, open floor with nothing in it, all of that looks near-identical to the matcher. A global fix needs a scan that could plausibly belong to one spot and nowhere else on the map. Easy test before you touch any config: park the bot somewhere that hasn't changed since you built the map. Same corner, no new furniture within lidar range, and let it run with no hint. If it locks on there, your launch setup is fine and the map is the culprit. If it still flops around even in the untouched area, then it's worth going back to the repo and checking whether the author ever ran it on real hardware rather than a simulation or a recorded dataset, because that difference shows up fast in exactly this way.
One more thing that bugs me about it. A proper global localizer shoud throw out several candidate positions from the first scan and sort them out over the next few seconds. If yours needs a manual estimate before it behaves, it's acting like a tracker, not a finder, no matter what the readme claims. If it turns out the config is fine, rebuild the map on a quiet day with the current layout and try again fresh. That alone fixes this exact symptom more often than any flag does.
Had the same flop-around behavior in my garage last winter. Three walls looked identical to the scanner and it just guessed forever. What actually fixed it for me was a scrap of pegboard I leaned against one wall at an angle, nothing fancy, just something with an odd profile that exists in exactly one spot. Once that was on the map, the thing locked on from cold starts every time. If your untouched corner test works but the rest of the shop doesn't, a couple of cheap wierd-shaped landmarks scattered around might save you a full remap.
Gotta push back on the pegboard idea. Sure it works, but now you've wired a fake landmark into your map, and the localizer treats that scrap of wood as ground truth forever. Shift it while cleaning the garage six months later and the thing either loses lock or worse, snaps to the wrong corner while acting totally confident. I went through that cycle twice with a stack of crates I left as a feature and swore off the trick. If a room needs decoys planted in it before the algorithm can find its own position, the problem is the algorithm, not the space. Has anyone tested how it handles a landmark that quietly disappears?
One thing I'd check before anything else is wheather that global mode is even switched on by default, because on a couple of stacks I've run the cold-start search lives behind a parameter that defaults to tracking-only, and the readme never mentions it. Mine just sat there quietly following its own garbage estimate and I burned a whole evening on maps before I found it. Also worth confirming your laser-to-base transform has the right mount rotation, since a flipped or slightly skewed static transform makes every scan land a few degrees off and the matcher never converges no matter how clean the map is. Did the author include any debug output showing candidate poses, even a simple rviz marker?
This might be a dumb thing but did you try just driving the bot around slowly after starting it? Mine did something similar in my basement, sat there confused for a while, but once I pushed it forward a few feet the estimates suddenly snapped into place. I figured it just needs a couple different views of the room to sort out where it is, like squinting at things before it clicks. Not exactly hands free but way better than typing in a spot every morning. Does yours stay frozen in that junk position even after moving around, or does it eventually settle somewhere wrong and stay there?
Update: it was the last thing mentioned here, the cold start flag. It was sitting on tracking only out of the box, so it never even tried a global search. Flipped it on and it locked on from my far corner within a couple seconds. Still remapping the area this weekend anyway since the moved shelves can't be helping. Thanks for the pointer.
Join the Discussion
Have something to add? Share your thoughts — no account required.
Didn't find your answer?
Ask the community — no account required