Funny - I've never ridden rollers or wanted to, but aware of their difficulty I mentally conceptualised this design some time ago.

Their explanation is different to how I would perceive them working though. I can understand how it could cause your wheels to slow down and the rollers to spin faster as you get nearer the edge, but not understanding how that makes you return to centre.

My thought is that either
a) the outward slope pushes the tire contact patch away from centre, which then leans you towards the centre and (through gyroscopic precession) creates a self-centering torque through the steering, or
b) the tyre on the angled slope is like a leaned tyre on a flat surface and (through the way the contact patch works) wants to roll in an arc in the direction of the lean, ie back up the slope.
<edit; further pondering>
or
c) The edge of the tire contact patch closer to the centre is moving faster across the surface of the roller, because the roller is a bigger diameter. Hence it creates more drag than the edge further away, and so there is a (centering) steering torque generated by asymmetric friction across the width of the contact patch.
</edit>

Random further thought for the day (yes I was pondering this earlier today after seeing the new Kurt trainer on video):
Basically trainer rocker plates seem like a crap idea because they don't replicate proper bike movement, the roll centres are all wrong and the steering is still fixed. So...
What if you attached a front wheel roller to a kurt trainer? That would make the front wheel want to wander, then the bike lean a bit, which you can then counteract naturally with steering input. Bascially it would be like rollers but with limitations to the range of movement, as well as proper resistance.

I've thought about that one also. The rock/roll is not the correct climbing dynamic.

My idea was to have the trainer bolted to a pair of heavy duty rails that are curved down toward center and have springs on the outer limits.

As you climb out of saddle shifting the weight back and forth, the bike would properly slide left/right like it does on a real climb or sprint. The entire assembly would be a maybe 18" x 36" rectangle. 36" the direction the bike sways left and right.

I've always thought the front doesn't move nearly as much as the back.

or,
d) The smaller diameter portions of the rollers create more rolling resistance due to geometric effects at the contact patch. The wheel tends to "center" because the rolling resistance is lower there. It's the same effect I compensate for when "translating" small roller Crr tests to an expected flat surface Crr.

So...the explanation in that video sounds like it's pretty much spot on.

---

http://bikeblather.blogspot.com/

But to me their explanation doesn’t explain what causes it to pull back to centre. Worse, if you’re leaning to the right and move right and the contact patch moves left, it will tilt you further to the right, accentuating the problem

Yeah...I hear what you're saying. I'll bet it's a bit of countersteering going on as well (i.e. turning the front wheel to the right causes the bike to lean left, etc.), but then it wouldn't make sense why they have one of the rear rollers crowned as well.

---

http://bikeblather.blogspot.com/

my guess b) would explain that.
Picture the bike off centre on the roller, and visualise the strip around the tyre that is contacting the roller. It's a slice of a cone. that cone is going to want to roll around the virtual tip of the cone, ie steer back to centre.

P.S. I googled "rolling a cone" and didn't get many helpful images...

Just checked their website and it seems that they went under. Anyone have the details?