Which one is better for importing assets with skeletal animation and why: COLLADA or FBX?

Neither and both. Evaluate the API interfaces and choose the one that seems most sane to you.

I'd really use Assimp for this. They're both a mess, and vary from the modeling packages you use to export them.. Assimp will make your life easier in every way and will probably increase the odds you get something reliable working.

That's just a collada problem primarily - and assimp won't make that problem disappear.

But what problem has COLLADA?I searched a liitle on google and i found that fbx is buggy.But COLLADA seems very clean and it has animation,skeletal animation,uv,materials etc.. And it has also a very good physics simulation integrated,with Bullet Physics(used also by GTA IV and V!!) and other engines...materials with friction etc....

but collada has the linear interpolation integrated...so you just make the animation in your 3d programme and import it to your game when it is needed...also fbx is very buggy and people who worked with fbx sdk are very unpleased...

But what problem has COLLADA?

There is only really one flaw. It attempts to write data into 'common profiles', and also allows 'custom profiles'. The theory is that anything you write into the common profile will work with all packages. In practice, what actually happens is that a lot of exporters simply dump data into a custom profile instead. This can lead to a number of if(max) elif(maya) elif(xsi) code constructs.

It's not really a deal killer though - it's just something you might encounter if you need to support mroe than one of: Max/XSI/Maya/Houdini.

FBX is the clear winner and for a very good reason: It supports inspecting a scene at any given time within an animation.

That argument is a fallacy. The question here is which data transfer protocol is most useful for the OP. Being able to evaluate animation and geometry deformations does not help you transfer data, it's just another feature that you won't be using within your game.

At run-time you want to perform only linear interpolation between key frames, as linear interpolation is the fastest.
In practice, there are many ways to perform interpolations between key frames, but using the FBX SDK you can create a
run-time system that only uses linear interpolation.

No. At the very minimum you want to use normalised linear interpolation for rotations, and linear interpolation for everything else.

Basically you run through an animation and set a bunch of key-frames, then you eliminate any key frames that don’t change the result if linearly interpolating between the 2 key-frames around it.

That's one of the worst compression schemes available for runtime animation systems. It fails to provide any random access capabilities at all, so you'll be forced to utilise precious memory bandwidth to find the correct keys. It simply won't scale.

The FBX SDK will generate all the values for each track considering each interpolation mode for you, so with the FBX SDK it is very easy to handle efficient animations and key-frame reduction.

As will Maya, Max, Xsi, Motionbuilder, Endorphin, Morpheme, Houdini, and even blender. That kind of processing is best left to the artists. Doing it at the asset packaging stage is only going to introduce animation accuracy glitches that will only appear within the game. Taking the control of anything away from the artists is always a bad idea imho.

I’ve worked with FBX for years, I have encountered no bugs, I am not displeased

Appeal to authority fallacy.

I've worked with FBX for years, and I think it's a horribly broken system that requires far too much hand holding from animators to make it work reliably.

By far the biggest problem with FBX is the excessive number of memory allocations the SDK utilises when performing interpolation internally (pretty much one alloc per float-key). It's just about ok to sample the keys from a single character, but if you have multiple characters & takes within an FBX, then you will have to start jumping through numerous hoops to keep the number of allocations within check (as in, you will be writing a custom memory allocator for the library, usually followed up by an extremely unhelpful chat with the Autodesk devs).

The second related problem, is that FBX has a notorious failure rate as soon as you start using secondary animation generated from expressions / constraints / IK rigs / etc. Now, if you've never seen it go wrong, you're ether not doing anything that complicated with it, or you're working with a bunch of artists who know how to correctly bake out their animations before export. If you are simply exporting characters from Maya or XSI, FBX will fail more often than not. I'd go so far as to say that most animation/rig import problems I saw when working on morpheme, were as a direct result of FBX.

Collada has pros and cons.
FBX has pros and cons.
dotXSI has pros and cons.

Evaluate them all yourself, and choose the one that works best for your project. They're all crap in one way or another, so don't spend too much time worrying about the right choice....

Rob,thanks for this opinion.I decided to use COLLADA,because i don't need anything but 3dsmax...:D. And,if you use more 3d programmes, you can import all your models to 3ds max and then export to *.dae ...also,3ds max will make the best rigging for you(biped),that's why i don't see why would you choose another programme for that...so,collada has more pros than the others,and(for me) no cons...