Blender has lots of options texture baking.
If one situation is showing up errors, there are multiple fall-backs.
The most obvious issue people have with Object Space maps is that Commercial game Engines rarely have full support if any at all for this format.
These maps can be usable outside of Blender if you:
(1) Know how to write GPU shaders
(2) Know how to send the proper rotation matrix to the shader.
That's a whole other topic.
In Blender, Object Space maps offer by far the easiest and fastest workflow available.
Any Blender tool which alters a model produces information that an Object Space map is able to capture with no special treatment needed.
Your high-Poly model can be the very same model as your low-poly model.
Only one asset is needed here. Often with Tangent Space baking, it will be necessary to trace a ray from one model through to the other using "Selected To Active"
Then you may also want to use a cage; Now that's 3 assets instead of only one to keep track of, and they also have dependencies with one another that can break the bake if anything is changed on one of them.
In comparison, to bake an Object space map using Blender internal is one of Blenders easiest tools to use.
It's only three or four easy steps to produce one.
Only one model is needed for this to work.
(i) Make a UV layer.

(ii) Click the 'New' Image Button or drag and drop a texture from your Hard-Disk-drive onto the newly created UV layout in the UV Image Editor. Be sure to select all the faces prior to adding the image texture.
(iii) Press the 'Bake' button in the Render panel.
The animated .gif makes it look bad but it should come out flawlessly so long as the UV's are nicely shaped and proportional.


(iv) Use the map in a render.

(v) Add a multi-resolution modifier and change to Sculpt Mode. Increase the Sub-Division iterations as high as you can go and start sculpting.

You can bake object space maps any time simply by pressing the 'Bake' button

That's all that's needed for a basic bake and it can be used for Blender Render(Internal), Cycles and GLSL renders.

It should also be mostly easy to do for the BGE shader system so long as the proper Reversed Rotation Matrix can be accessed.
The efficiency gain here is really substantial both in terms of memory requirements and also with the GPU shader instructions.
(i) In the Vertex shader, an Object Space map shader requires only a Simple 3x3 Matrix multiply against the light position. No Normals, Tangents, BiTangent's are needed so the arrays can be much smaller.(less memory, shorter transfer times, more of the model fits into the faster but small cached memory pool which helps to eliminate redundant calculations :)
The efficiency of object space maps leaves the GPU free to do other lighting effects.
A limitation of them is that you should never "Apply The Rotation" of your models because that extra Rotation info from Object mode is being used in the shader to correct the object space map vectors. That's a tiny limitation, in practice it's not usually a big deal so long as you are aware of it.
Scatter and Rotate small rocks and whatever else all over all the scene and the normal maps should hold up fine just as Tangent maps do.
You can Join and Separate Objects and still have valid Object space maps so long as you do not rotate any of the meshes in Edit Mode before re-separating them.
The Displacement Modifier also bakes to Object space maps with results the same as with Multi-res Sculpting. It works beautifully. :)
Blender has excellent support for Object space maps even when they are being rotated so long as the model is not an animated character mesh.
I don't think the bone system has been upgraded to support them yet, but for mechanical toy-like models with hinges and for background objects or still renders, they are ideal due to the simple setup and the ray-trace free workflow.
Here is an example of a particle system that uses duplicated objects which are set to rotate across the surface of another model.
Cycles is handling the normal map rotations without any issue that I can see.
The sphere that is being duplicated is very low poly but appears to have a lot of detail.
The 3D viewport response is excellent, even with many more instances, since there are no high-resolution generative modifiers being used.
