Replies to This Discussion

Permalink Reply by Caleb Chamberlain on August 4, 2009 at 7:58am

Yes, of course you can combine a variety of sensors using any number of algorithm, perhaps most commonly a KF. What I am saying is that near hover the accels can't give us much information that can be used to correct gyro drift: we can only use them to get the magnitude of the thrust vector. Provided that there are body accelerations that arise independently of thrust and attitude (ie. from wind disturbances), we can also get some information about lateral acceleration. But that doesn't help very much if you want your accels to contribute to long-term drift correction.

It doesn't matter how you phrase the problem mathematically; if the accels can't physically measure something, you can't change their nature by throwing equations at them.

Permalink Reply by Caleb Chamberlain on August 4, 2009 at 8:35am

Your research, BTW, sounds really interesting. Have you published anything on it yet? I would be interested to read it.

Permalink Reply by bGatti on August 5, 2009 at 5:45pm

A bit rich on jargon, but interesting:

I think what you're saying is that, like gravity, the sum of polar magnetism is largely constant, and with a multi-axis magnetometer, you should be able to detect the attitudinal (non-compass-axis) angle of the polar magnetic field.

... and using the absolute attitude (magnetic attitude), confirm that a neutral accelerometer is also a flat position magnetically - rather than a sideways drifting state?

Permalink Reply by Roy Brewer on August 14, 2009 at 7:00am

Caleb,

Did you see this paper?

Autonomous Navigation and Exploration of a Quadrotor Helicopter in GPS-denied Indoor Environments, Markus Achtelik, Abraham Bachrach, Ruijie He, Samuel Prentice and Nicholas Roy

(found it using Google Scholar)

- Roy

• ▶ Reply

Permalink Reply by Caleb Chamberlain on August 14, 2009 at 7:34am

Roy,

Yeah, we've had our eyes on these guys; they just won the International Aerial Robotics Competition (IARC) with the setup they describe in this paper.

We are a little behind them... we made the mistake of designing our own quadrotor, so instead of focusing on algorithms, we focused on hardware design. We should have bought one like they did.

• ▶ Reply

Permalink Reply by Xander on August 1, 2009 at 2:18am

Damn you nature! Always causing problems...

You're right. It's counter-intuitive (at least for me), but the accels do always cancel out. At least so long as you're maintaining altitude.

Odd that so many projects use accels and no one I've read about has noticed that it shows nothing. Makes me feel like we're still missing something, but I'll admit your logic does add up.

• ▶ Reply

Permalink Reply by Caleb Chamberlain on August 1, 2009 at 7:37am

In most cases, I think that the upward force will not always be m*g; even if you have an altitude hold loop running, it may not respond quickly enough to update the force vector in the event of a disturbance. The effect would be that the accels measure SOMETHING, but not the correct angle (it would be much too small), and in the long run, still centered around zero.

One possible reason that nobody notices the accels aren't working is because the estimated states seem fairly accurate. With GPS position hold in-the-loop, you wouldn't even notice there was a problem. And without GPS position-hold, the angles would seem fairly close, since you are only slowly pulling the angle estimates to zero, and on average they are often close to zero anyway.

I have to admit that most of the research I do is in mathematics, so I haven't personally evaluated this with real hardware. I heard second-hand that a guy at the Field Robotics Center in Pittsburgh set up a test-bed and proved that accels don't work (in case the basic math was missing something). I don't know who it was, though. It is possible that we are missing something, and that in the real world there is something different that allows them to work. I guess it would be easy enough to test; simply fly a rotorcraft near hover using gyros for rate stabilization, and log all the accelerometer data and gyro data.

I have an Ascending Tech Hummingbird (same airframe you posted the link to for demonstrating maneuverability); in angle-hold mode, it does start to drift after a while, and I always assumed it was because they are trying to use accels when they don't work. Maybe I'll do some datalogging to quantify it more carefully.

• ▶ Reply

Permalink Reply by Xander on August 1, 2009 at 2:14pm

The only way I could see the accels actually functioning is if we're reaching steady state a lot sooner than you think. Experimentally, giving a quad a 5 degree tilt seems to have it going at a pretty steady speed after the initial acceleration. I do realize how the forces involved work but perhaps since only a small portion of the thrust is going to lateral movement, and there's so much turbulence it can steady out quite low.

Just a thought...

• ▶ Reply

Permalink Reply by Caleb Chamberlain on August 1, 2009 at 2:23pm

That could be. Maybe I'll do some testing and see what happens.

• ▶ Reply

Permalink Reply by Roy Brewer on August 4, 2009 at 4:59am

Caleb

Can you please post those papers that go into detail about this? Don't get me wrong - I believe you completely! But, all the academic papers I've seen on quads or other hovering vehicles seem to imply the accelerometers give a "gravity vector", which we know is wrong. If it helps, I have access to AIAA and IEEE sites, so just some notion of title and/or author will suffice.

The way I keep it straight in my head is to think of the accelerometers ignoring gravity completely and just measuring the (specific) reaction force. For a quad, the reaction force (ignoring drag) is always in the direction of the rotor thrust - ie the Z axis. People may be confused further if they mount their sensors to a stand firmly connected to the ground during development and testing, since in this case the X and/or Y accels will be non-zero if the platform is tilted. However, the "reaction force" viewpoint still holds.

BTW, I'm a helicopter engineer by day, and an aspiring quad builder by night. I know our full size helis use accelerometers in their IMUs for position hold. Granted, they are GPS aided but they work OK without GPS - and of course they are much more accurate than the MEMS chips that we are using in our quads. But I'm not sure how they are able to get it to work. I wouldn't think helis are any different from quads in this regard...?

I'm also curious about your research, if you are willing to share...

- Roy

• ▶ Reply

Permalink Reply by Caleb Chamberlain on August 4, 2009 at 8:27am

Roy,

Earlier, I was too quick to say "accels are basically worthless." They are "worthless" in the sense that they can't really be used in the way everyone seems to use them. But they CAN (as you mentioned) measure the magnitude of the thrust vector, which is useful, and they can detect accelerations that are caused by external disturbances like wind and prop wash.

Helis might be a little different because the rotor head can move independently of the body, which a quadrotor can't do. So the body can sort of "hang" on the rotor head, which means you could measure the gravity vector in some cases.

I guess the whole point in my bringing up this discussion is that you have to be really careful about the assumptions you make. Most people don't realize that accels aren't doing what they expect.

I will ask my advisor about the papers; I am pretty sure they exist... At least, hopefully they do, after I spoke so confidently about them! :-)

• ▶ Reply

Permalink Reply by Caleb Chamberlain on August 4, 2009 at 10:43am

It looks like we have a lot of material on how accels behave here where I am studying (Brigham Young University), but we haven't published any of it... so I guess I was wrong when I assumed papers had been published on it. I made the assumption because among my colleagues here at BYU, and also at the Field Robotics Center at CMU (I worked there for a while), it seemed like everyone already understood what the accels were doing. Sort of a "common knowledge" phenomenon; though, maybe it isn't as common as I thought.

Since it hasn't been done, maybe we will put together a paper that explains what happens, and also explains why people seem to make it work anyway; it might be enough of a contribution to be publishable...

• ▶ Reply

• ‹ Previous
• 1
• 2
• 3
• Next ›
• Page