PDA

View Full Version : Any experience with Bumblebee2 stereo camera?



Pi Robot
08-26-2010, 09:10 AM
Does anyone out there have any experience with the Bumblebee2 stereo camera from Point Grey Research? (http://www.ptgrey.com/products/bumblebee2/index.asp). I may have an opportunity to add one of these to Pi Robot but they ain't cheap so I'm curious to hear if others have used them and, if so, with what operating system(s) and programming languages.

Thanks!
patrick

lnxfergy
08-26-2010, 09:46 AM
Since I know ROS is in your eventual plans, I've not actually seen many people (anyone?) using the Point Grey stereo cams under ROS -- but many people are using the Videre STOC (which is well supported under Linux and ROS).

A stereo camera is actually next on my list of big buys too (I just got a URG scanning laser though, so it's a ways off).

-Fergs

Pi Robot
08-26-2010, 10:10 AM
Dang, I don't know how I blanked on Videre so thanks for the reminder. Looks like they also have some nice mobile bases which I could also use depending on cost. As it turns out, they're just down the street from me but it looks like I need to fax in a request for a quote so I'll do that next.

--patrick

DresnerRobotics
08-26-2010, 11:44 AM
Their mobile platforms be mighty spendy...

zoomkat
08-26-2010, 12:03 PM
Natures warning sign: "Our representatives are pleased to provide you with a customized price quote based on your specific product interests and requirements". Either so expensive the price will cause an immediate heart attack, or they only make them when there is an order (aka, they've probably have only made a few).

Pi Robot
08-26-2010, 12:10 PM
Yeah, I've got my nitroglycerine tablets standing by...

lnxfergy
08-26-2010, 12:37 PM
Natures warning sign: "Our representatives are pleased to provide you with a customized price quote based on your specific product interests and requirements". Either so expensive the price will cause an immediate heart attack, or they only make them when there is an order (aka, they've probably have only made a few).

Or (as in this case) they are highly custom (the bases, hell, even the cameras, there's like 3-dozen different configurations). Videre is aimed at researchers -- researchers expect to hand in a very custom list of requirements, and get a fully assembled & tested platform back.

-Fergs

Pi Robot
08-30-2010, 08:46 PM
A stereo camera is actually next on my list of big buys too (I just got a URG scanning laser though, so it's a ways off).

-Fergs

Which URG scanner did you pick up and have you had a chance to try it yet?

--patrick

lnxfergy
08-31-2010, 04:30 PM
Which URG scanner did you pick up and have you had a chance to try it yet?

--patrick

It's a Hokuyo URG-04LX-UG01 (http://acroname.com/robotics/parts/R325-URG-04LX-UG01.html). Seemless integration with ROS using the hokuyo_node package. I had it broadcasting laser scans in about 10 minutes. Another 30-40 minutes, and I had a map of our lab (5 minutes to find double sided sticky tape, 2 minutes to debate whether attaching said tape to a $1100 lidar was a good idea, 3 minutes to attach the lidar/netbook/etc, 10 minutes to find out that the motor had died on one of the lab creates, rinse/repeat, 15 minutes of driving around collecting a rosbag of laser/odom data, and a few minutes to create a map with gmapping):

http://forums.trossenrobotics.com/gallery/files/1/7/6/8/map-ils.png

-Fergs

Pi Robot
08-31-2010, 04:45 PM
OK, that is just too cool. Time to put one of those on my wish list.

--patrick

lnxfergy
08-31-2010, 04:51 PM
OK, that is just too cool. Time to put one of those on my wish list.

--patrick

It's one of those things I wish I had bought about 2 years ago.

-Fergs

RobotAtlas
09-04-2010, 07:47 PM
My wish is they were at least 5 times cheaper. Better 10 times. :)
Until that happens, only labs can really afford them and not normal people.

How long do you thnk until that happens?

lnxfergy
09-04-2010, 09:10 PM
My wish is they were at least 5 times cheaper. Better 10 times. :)
Until that happens, only labs can really afford them and not normal people.

How long do you thnk until that happens?

Well, it took about 15 years for the cheapest model to go from about 10k to 1k. I'd guess the next 10-fold reduction will take even longer (as you can only make something *so* cheap before it gets really tough).

As it is, an $1175 laser has limitations compared to a more common $4-5k model. These include density of the scan, rate of scanning, and the valid range (this lower cost model has a blind window out to about 60cm, and tops out around 5m whereas 20cm->20m would be the range of a more expensive model, and that extra range is quite useful in scan matching for localization). A $100 laser is certainly going to have limitations compared to a $1175 model. We might go so far as to say that the PML is your $100 scanning laser -- as it offers a similar valid range to the URG-04LX-UG01, however with about 1/10 of the scan rate, and about 1/10 of the scan density -- at which point it's good for obstacle avoidance, but is pretty much useless for localization.

However, I think the question you have to ask yourself is: how much does the bot already cost, and what do you intend to do with it? I think that overall, $1175 for a scanning laser really isn't bad if you're intending to make a robot that operates in typical human-scale environments. For instance, you're already spending about half that amount on the PC-brain for any bot that can actually make use of a scanner. Typically, the drivetrain/mobility platform is several hundred more. Want to add arms to do something in that environment -- add a few hundred more. It quickly adds up. What sort of alternatives are there to a laser? Pretty much no reasonable ones. A 16-sonar ring like you might find on some research bots of the 90s would today cost at least 16x$25=$400, and I'd say that a single laser is at least 3x better than a sonar ring for the application of SLAM, particularly because most (all?) openly available SLAM packages today target lasers. A stereo-camera head could replace a laser -- but of course wouldn't have the wide field of view of a laser, and you really can't have the robot be moving too fast if you're using cheap webcams (the stereo matching will get all skewed), and a research-grade stereo setup is going to run more than this laser.

Finally, I don't think it's even the cost of the laser that makes it such that "normal people" can't access it (and are you saying I'm not normal? I own a personal lidar...), but rather what a lidar does. Most hobbyists aren't creating fully autonomous robots that roam freely in unaltered, human-scale environments. Even if they can afford a laser, do they even have a use for it? Do they use a software package, such as ROS, that provides utilities for using a lidar? And do they know enough to effectively use it -- even though the ROS navigation stack is an amazing step forward, it's still not plug and play -- tuning it to run effectively on a new robot platform is definitely non-trivial.

-Fergs

Pi Robot
09-04-2010, 09:54 PM
...Even if they can afford a laser, do they even have a use for it? Do they use a software package, such as ROS, that provides utilities for using a lidar? And do they know enough to effectively use it -- even though the ROS navigation stack is an amazing step forward, it's still not plug and play -- tuning it to run effectively on a new robot platform is definitely non-trivial.

-Fergs

This is a great point. I am 5 years into building Pi Robot (through many design changes, programming languages, libraries, etc.) and even now it's not clear to me how much I would use lidar even if someone bought me one for my birthday (one can hope right?) Having said that, one possible less expensive option is to get one of the new Neato vacuum cleaners and see if you can hack into their lidar--at $400 you get wheels, motors, batteries, cliff sensors and a lidar. See:

http://www.neatorobotics.com/

--patrick

lnxfergy
09-04-2010, 10:16 PM
This is a great point. I am 5 years into building Pi Robot (through many design changes, programming languages, libraries, etc.) and even now it's not clear to me how much I would use lidar even if someone bought me one for my birthday (one can hope right?) Having said that, one possible less expensive option is to get one of the new Neato vacuum cleaners and see if you can hack into their lidar--at $400 you get wheels, motors, batteries, cliff sensors and a lidar. See:

http://www.neatorobotics.com/

--patrick

The neato laser is described in an IEEE paper from 2008. They made a LOT of tradeoffs -- in particular, the whole sensor head spins. This presents an interesting issue: the slipring used for power is only rated for about 1000hrs of continuous operation (which is 3yrs at 1hr/day for a vacuum robot -- but only a couple of months at 4-6hrs/day if you want the bot to be on more often). It's also only getting 1 deg resolution (about 1/3 of the low-cost URG, but scans full 360 degrees instead of just forward 180). The valid range and scan rate is about the same as the URG.

I'm guessing hacking the laser would be tough -- in their paper, the data is sent wirelessly from the spinning sensor head back to the bot -- so you'd probably have to hack the wireless aspect off the main board.

-Fergs

RobotAtlas
09-05-2010, 11:19 AM
Well, it took about 15 years for the cheapest model to go from about 10k to 1k. I'd guess the next 10-fold reduction will take even longer (as you can only make something *so* cheap before it gets really tough).

It does not have to be any particulare technology. I was mostly talking about a sensor that will allow doing SLAM well. Without SLAM, a robot is still not very different from RC car (except RC car doesn't fall :) ).

Considering we already have Kinect and Neato, which both have reasonable cost, I bet it will be 2-3 years before affordability in these sensors happens.


(and are you saying I'm not normal? I own a personal lidar...)
Wow, I'm ipressed. I thought your lab bought one, not you personally.
Considering how much time you spend doing this stuff, I'm totally with you on justification for that laser. It's definitely worth it and as you said, there's nothing cheper in that class right now.

The "normal" people I was talking about are the ones who spend $4,000 to buy their car. :)

RobotAtlas
09-05-2010, 11:24 AM
The neato laser is described in an IEEE paper from 2008. They made a LOT of tradeoffs -- in particular, the whole sensor head spins. This presents an interesting issue: the slipring used for power is only rated for about 1000hrs of continuous operation (which is 3yrs at 1hr/day for a vacuum robot

I can already see how in 3 years we will be able to pick up those Neato's on ebay for $20.
Hopefully it will be easy to replace those sliprings too.

lnxfergy
09-05-2010, 12:40 PM
Considering we already have Kinect and Neato, which both have reasonable cost, I bet it will be 2-3 years before affordability in these sensors happens.

I'm a bit more skeptical.

The neato laser has potential -- but I think that the longevity of the sensor, and the density of the scan are both issues that will have to be overcome -- I doubt people are going to rush out to buy even a $200 laser if the estimated lifespan is 8 months (at 4hrs per day) or only 3 months if you leave it on all day.

I'm really not sure about the Kinect. I have a feeling that it, like many other non-traditional game controllers, is going to come up a bit short -- for instance, the Wii has never really caught on with serious gamers, because it's just not robust enough to adequately capture the desired input data. I really don't see the Kinect catching on as a SLAM sensor because the range is going to be limited, the data is likely to be quite sparse, and I have a feeling it's going to have a real problem with motion (yes, the stereo setup is designed to capture motion, but within a non-moving environment. If you slap the sensor on a moving platform, it's a different game). It might make a fine robot-input device for people interacting with the robot. For $150, it might give some data, but it's probably not going to work with existing openly-available SLAM packages. I'm actually more hopeful about it's microphone array than it's use as a 3D camera.

-Fergs

SK.
09-05-2010, 02:21 PM
Well, I'm not quite as sceptical. There already are a bunch of people working on RGB-D cameras and exploiting their favorable properties for SLAM and other interesting problems. For current state of the art examples see http://www.seattle.intel-research.net/RGBD/RGBD-RSS2010/

Especially note the "RGB-D Mapping: Using Depth Cameras for Dense 3D Modeling of Indoor Environments" presentation which shows extremely nice results (probably using a sensor from the Primesense line, at least that's what I gathered).
So let's wait and see. I wouldn't be surprised if cheap RGB-D sensors cause a small revolution in (indoor) environment sensing in the next few years.

lnxfergy
09-05-2010, 02:38 PM
Well, I'm not quite as sceptical. There already are a bunch of people working on RGB-D cameras and exploiting their favorable properties for SLAM and other interesting problems. For current state of the art examples see http://www.seattle.intel-research.net/RGBD/RGBD-RSS2010/
...
I wouldn't be surprised if cheap RGB-D sensors cause a small revolution in (indoor) environment sensing in the next few years.

I guess I should be a bit more clear here -- I'm skeptical about ultra-low-cost sensors. I'm not saying the technology is impossible -- I'm saying it's not super cheap. The problem is: what exactly is "cheap"?

There is a lot of research into the necessary algorithms, hardware, etc. The problem is: on a mobile platform, those sensors are still quite expensive. There's a huge difference in performance between a pair of webcams and a Videre stereo camera -- especially under motion -- due to how the physical camera captures a scene. High quality cameras really aren't coming down in price lately, as manufacturers have done a pretty great job of reducing the production costs as far as possible -- instead, we're simply seeing the feature set on "low-end" cameras go up (but the "low-end" price is well above what RobotNV is asking for).

So, if you consider "cheap" to be $1000 for a primary sensor -- it's almost there, but if you insist that cheap is $50-200, you're going to be waiting a very long time (currently, sensors like an 8-pixel thermopile costs about $120). Something like an RGB-D camera is easily the core sensor of a very advanced robot, heck, if I could get a high quality RGB-D cam with a reasonable range for about $1500, I'd jump on it (as it is, I'll be trying to pick up a real stereo-pair later this fall).

-Fergs

SK.
09-05-2010, 04:06 PM
High quality cameras really aren't coming down in price lately, as manufacturers have done a pretty great job of reducing the production costs as far as possible -- instead, we're simply seeing the feature set on "low-end" cameras go up (but the "low-end" price is well above what RobotNV is asking for).
Well I'd hazard the guess that all research-level stereo cam systems, Hokuyo LIDARs etc. are still "handmade" for at least parts of production, as the numbers made are probably a few orders of magnitude lower than those of a gaming system that Microsoft introduces worldwide. A representative of a ToF-cam producing company I won't name here also told me that the high price is basically solely to "handmaking" the things, and that mass production will likely bring down the cost immensly.


it's almost there, but if you insist that cheap is $50-200, you're going to be waiting a very long time (currently, sensors like an 8-pixel thermopile costs about $120).
I (and a lot of people working in the field I talked to) think that's likely a litte on the pessimistic side.To cite the guys from the University of Washington in their paper: "Very soon, small, high-quality RGB-D cameras developed for computer gaming and home entertainment applications will become available at cost below $100.". Also note that coincidentally, they're from the same place where Microsoft's HQ is and mention the Primesense sensor in their paper. So maybe they know a little more than you and me ;)
But no use in arguing, let's wait a year or two and some trend will probably be visible.

Adrenalynn
09-05-2010, 04:14 PM
Alas, you missed Ferg's point that the low-end gaming cameras are designed to be far sparser than we require. Raising the density actually hurts them.

Of course, mass production always drives prices waaaay down. But there needs to be correlation to purpose.

RobotAtlas
09-05-2010, 04:38 PM
Those SwissRanger 4000 sensors used in RGB-D are around $10,000, right?

SK.
09-05-2010, 04:45 PM
Alas, you missed Ferg's point that the low-end gaming cameras are designed to be far sparser than we require. Raising the density actually hurts them.
Well if Wikipedia is right (yeah, I know.. ;) ) you can get 320x240x30 distance samples from the camera per second, that's more than 50 times the number of data points you get from a Hokuyo UTM-30LX. Of course, with severe range limititations, sensitivity to motion and natural light etc. Still, if you're intelligent about it (like the guys from Dieter Fox's lab) you can do some pretty nifty stuff with the data, combining methods from Visual SLAM with ICP registration only of selected parts of the point clouds and all that.
The amount of data from pretty much any laser scanner looks sparse compared to what's coming in from RGB-D cams (likely even the cheap ones). I'm of course well aware that the quality of distance samples won't be able to compare to those of LIDARs.
Two months and we'll know a little more. Will be interesting if (or when) Kinect can be hacked and how good the sensor really is.

SK.
09-05-2010, 04:54 PM
Those SwissRanger 4000 sensors used in RGB-D are around $10,000, right?
RGB-D refers to any system providing RGB and depth data for pixels. So that includes ToF systems like the SwissRanger4K or CamCube, as well as structured light based ones like the Primesense one used in Project Kinect.
From what I heard the SwissRanger 4K should be around 6000$, not completely sure though.

Here is a good overview:
http://www.hizook.com/blog/2010/03/28/low-cost-depth-cameras-aka-ranging-cameras-or-rgb-d-cameras-emerge-2010
(and there 9K$ is given for the SR4K)

billyzelsnack
09-05-2010, 05:47 PM
RGBD is an obvious candidate to become a cheap commodity. I'd wager it'll be less than 2 years before we even see one integrated into the iphone.

RobotAtlas
09-15-2010, 08:37 PM
I found a video talking about Neato and its SLAM system
http://www.parc.com/event/1080/neato-xv-11.html

The video is 55 minutes.

Another short video of its SLAM system is here:

http://www.youtube.com/watch?v=zodC8EFvh7g

Just keep in mind the whole thing is $400 brand new.

zoomkat
11-12-2010, 10:52 AM
That is the video with the "feet" walking around. Just wonder how the bot knows how wide the feet are. I don't think lasers have the ability to detect the back edge of things.

RobotAtlas
11-12-2010, 11:19 AM
That is the video with the "feet" walking around. Just wonder how the bot knows how wide the feet are. I don't think lasers have the ability to detect the back edge of things.

It's a great question to ask Neato. My only guess is they are making some assumptions.
It doesn't make sence to mark everything behind a new obstacle as occupied.
So I think they use a default thickness for dynamic obstacles.

lnxfergy
11-12-2010, 12:25 PM
Looks like a typical occupancy grid approach -- so then the system doesn't care about the thickness of obstacles, they simply mark each cell as occupied or not. The cells corresponding to leg fronts will get marked as obstacles when a person is there, and then cleared out again when the person walks away (by a laser scan that goes through them to the original wall).

If, on the other hand, you're looking into how you can find legs in a laser scan, that's a slightly different probably, typically accomplished using a machine learning approach. (See Arraz, Mozos and Burgard, 2007 and 2008 for a very well designed Boosted method for such a leg classifier).

-Fergs

RobotAtlas
11-12-2010, 12:53 PM
Looks like a typical occupancy grid approach -- so then the system doesn't care about the thickness of obstacles, they simply mark each cell as occupied or not. The cells corresponding to leg fronts will get marked as obstacles
So do they mark only one cell deep (depth from the point of view of robot)? The question is how do they know about feet _back_ sides?

lnxfergy
11-12-2010, 12:59 PM
So do they mark only one cell deep (depth from the point of view of robot)? The question is how do they know about feet _back_ sides?

Yes, the endpoint of a lidar beam falls in a single cell and that is all they mark as occupied, all other cells on the path from the lidar to the end of the beam are marked as not occupied. As for seeing the back side of the legs, they don't. If you look at the video, legs appear has half-circles (or crescent shaped arcs).

From a navigation planning perspective, you only care about the side of the legs that face you, because in order to collide with the back side of the leg, you'd have to go through the front side (or go around it, at which point you'd see a different perspective).

-Fergs