Below is the original proof of concept chassis as I built it. I hacked a tiny piece of test code from the QuAD stuff and away we went. I won't post such simple pieces of nqc code again, but something like this, or the equivalent in another RCX programming mechanism, is something you'll need over and over, if you're going to work out the mechanics. File it away for next time.

I put the thing on the track, it stuttered for a couple of feet, and jammed. Attempting to free it by jiggling it with my finger resulted in the motor popping and falling off, dislodging the RCX. Fortunately, although I test stuff live, I also prepare for the worst so they plummeted about six inches onto the carpet. Hint: if you have a dark multi-coloured carpet, put a white sheet down – it makes finding the small black bits so much simpler.

I re-assembled it without the motor and ran it up and down with my finger. The timber wasn't planed quite as accurately as I had thought and the unit bound in a few places. I could have sanded the high spots down (and I will, soon), but this kind of thing is also normally indicative of being too close to the limits of tolerance.

I wanted to increase the gear ratio, but to do that, I had to find somewhere to put another shaft. Under the existing shaft looked like a good bet, but the geometry was against me. The solution was to use the same size of tyres at both ends. I could then remove the shim at the rear, raising the frame sufficiently to get another axle in. The geometry became simpler and I could mount the motor above the wheel rather than in front, shortening the chassis, without raising the C of G far enough to worry about the balance.

As Fell found out some time in the 1870s, if you get the geometry right you need remarkably little in the way of guidance to stop it falling off. His later designs use much less in the way of stabilisers than the Aldershot test track did. Usually, when a design change makes things simpler as well as better, it's right. Ria also started to look like something with which I was familiar: a MotorRail motorised skip from the 1950s. These were designed to work automatically in confined spaces, running along moveable rails that looked a little like tall crash barriers: the beauty of a monobeam system is that at slow speeds tracklaying doesn't need to be too accurate if the underlying engineering is sound.

Heartened, I fastened the RCX in place, making the assembly more rigid. For good luck I added the driver and ran it up and down the beam. Now with 12:1 reduction, it ran at about a scale 15mph (from the size of the driver). Better, it didn't bind any more, having more power, a shorter chassis and less slop. Repeated runs showed that unguided positioning was within a half inch in six feet. Not brilliant, presumably because of wheelslip at the points where it bound, but better than QuAD. This means that the ball-sorting code should work unchanged if I can mount the sensors to match. Both re-usable code and software influencing the hardware design. Mmm, that tastes so good.

I just had to go in the next morning and run it some more while I checked the email. I missed my bus, meaning that I missed my train, but it was still worth it. I expect more butchery as I add the grab and the sensors, but I think I have a solid design. It turns out to be simpler than QuAD, but that's because a lot more thought has gone into it. I'll photograph the build-up next, and then work on the software for a bit.