Hacker News Reader

Comment by georgecmu | Hacker News Reader

He's subscribing to Omnistar HP service. Basically, it's a float RTK solution relative to a synthetic base station. They boast 10 cm error 95% of the time. It's expensive (IMHO) and breaks whenever sun flares up.

True RTK + IMU positioning systems will get you 1-2 cm 95% of the time (see Applanix, Novatel).

View on HN

Interesting! I don't use this stuff in anger, so my questions are merely idle curiosity.

Omnistar refers to the HP accuracy as "real-time"[1] and:

> [it] will usually have a 2-sigma (95%) horizontal error of about 6 centimeters and a 99% horizontal error of less than 10 centimeters

Do you happen to know what kind of movement this is based on? For example, if it was attached to an RC car (assuming that the equipment is this compact, it may not be?) would you still be able to get this level of precision? Or is it designed for more consistent movement patterns?

[1] http://www.omnistar.com/about.html

My guess is that it doesn't assume anything about your motion model. The real question you should ask is what's the frequency of the positioning solution. Your generic GPS unit computes solutions at 1Hz, more sophisticated ones do it at 5Hz, surveyor quality units can do it at 10-20Hz depending on firmware, but I think they're pushing the limit and probably assume something about the motion model. Applanix and Novatel solutions that use inertial sensors definitely make pretty strong assumptions of the motion model (see Kalman filter).

OmniSTAR in particular and DGPS in general will operate under assumption that errors due to ionospheric and tropospheric effects change slowly in time and space. If this assumption is correct, then dynamics of the vehicle won't affect the corrections themselves.

My point with all this is that 99% of the corrected GPS solution should theoretically be within x cm of some point on the true path of the vehicle no matter how it moves in a plane. The question is, if you're interested in real-time solution, how far will the vehicle be from that point by the time you get this fix, and if you're interested in capturing vehicle's travel, how well will the curve fit through these points match the true vehicle path.

Exactly what you said. But OmniStar is giving it away to us for free so why the heck not! :)

Just wait till your robot drives under a tree canopy )

There are other ways to get even greater accuracy than an applanix. But they require pre-indexing the surroundings and using a lidar to position yourself within the "virtual" world.

This would require lidar error to be less than 1cm. If you know a sensor like that, I would appreciate a link.

The riegl for sure. Others have made the velodyne work.

Haven't worked with Riegl, but I can tell you that Velodyne isn't even close. Even their official datasheets state range error of 5 cm; reality is much worse.

You probably do some sort of filtering on the raw input, maybe some sort of landmark based SLAM.

The factory calibration on the 64 diodes of the velodyne is bad. It needs to be fixed. There is a paper out of Stanford that tells you how to fix it. I can't think of the name off the top of my head but if you email me I can dig it up.

Also the error grows with distance. So if your bot is navigating off of objects 200 meters away, yeah you'll be more off. But you can use points closer to it and you'll get better results. Look around and you'll see that the velodyne has been made to work. Nobody loves it though.

The riegl has very low beam divergence, is beautiful, survey accurate, not waterproof and $300k+.

Don't forget the workhorse SICK scanners. They are cheap and have their pros and cons.

Reply on news.ycombinator.com