How accurate must my odometer reading be for SLAM ?
I am writing this extra section because it says my question body does not meet the quality standard.
1
-
$\begingroup$ This is very specific to your exact sensors and system. Many SLAM implementations don't use odometry at all. If you fully detail your system it might help get responses, but I would not expect much. $\endgroup$– hauptmechMar 14, 2016 at 2:57
Add a comment
|
2 Answers
$\begingroup$
$\endgroup$
4
That's what's great about SLAM: it doesn't matter. As long as your sensor noise is modeled appropriately and the odometer readings can be translated to position changes, the general algorithm holds.
In a practical sense, with very high uncertainty from odometery, it will update the position just fine, but will seem to ignore the odometer when any other information is available (e.g., laser and static landmarks). Without other information, your maps may look like junk, but it will faithfully produce an estimate with correspondingly high uncertainty.
answered Mar 14, 2016 at 14:23
-
$\begingroup$ I have a follow up question: I a have the Lidar and odometer data from a test run. If i put my starting position at [0,0,90] it gives bad map. However if i shift it away from [0,0] to something like [50,50,90] the map seems fine.How can this be ? Thanks again $\endgroup$– J.ChewMar 14, 2016 at 15:43
-
$\begingroup$ Follow-up questions need to be posted as new questions. Otherwise it's very difficult to answer them properly as they do not get the proper visibility. Thanks! $\endgroup$ Mar 14, 2016 at 15:44
-
$\begingroup$ "As long as your sensor is unbiased and has anything close to Gaussian noise" this assumption is unnecessary. This relates to the filter not to SLAM since other filters like particle filter work well without the Gaussian assumption. $\endgroup$– CroCoMar 17, 2016 at 15:35
-
$\begingroup$ @CroCo Absolutely correct. Updated the answer. $\endgroup$ Mar 17, 2016 at 17:08
$\begingroup$
$\endgroup$
How accurate must my odometer reading be for SLAM ?
There is no specific level for the accuracy. This is why the modern approaches for building SLAM are probabilistic approaches. This is due to the fact that noise in any sensor is inherited even with very accurate ones. The heavy load with very noisy sensors will be on your filters. You need very good filters to at least run SLAM in very simple maps (e.g. a bunch of points).
