It may look more like a golf cart than the golfers that sit in them, but this robot is actually pretty handy at putting. The handwork of researchers at Paderborn University in Germany, Golfi – as good a nickname as any, I guess – uses neural networks to figure out the optimal way to line up a putt, and how hard to hit the ball for putting perfection.

Well, not quite putting perfection. The researchers say it completes a successful putt between 60 and 70% of the time, but that’s still better than golfers beyond the pro circuit.

And the imperfect performance probably isn’t helped by the fact that Golfi will sometimes roll over the ball, knocking it out of position. This would once have incurred a one-stroke penalty, but this punishment has now been removed according to the USGA. Lucky Golfi, though it will still have to replace the ball, and then do all its calculations afresh.

These calculations begin with one of the many Xbox Kinect cameras made redundant after Microsoft finally conceded that people didn’t want Kinect on their Xboxes. The camera takes a depth map of the green, which is then fed into a physics based model which factors in things like the green’s rolling resistance, the weight of the ball and it’s starting velocity, before simulating 3,000 random shots.

Golfi then trundles over to the ball, and uses a belt-driven gear shaft to move a putter with the (hopefully) correct power and angle. All of this happens in just five minutes, which wouldn’t make it a great golfing buddy, but is still a vast improvement on feeding the system real-life shots, which would take 30 to 40 hours and definitely see you getting your club membership revoked.

That assumes you can actually take Golfi out on a real golf course, of course, and you can’t. Even if you did persuade the front desk that you’d just brought your own rudimentary buggy for the day, Golfi wouldn’t work in the real world as the Kinect currently takes it’s scan of the putting green from the ceiling above an artificial course. Short of attaching it to a drone, there’s no way of replicating this in the real world for now.

That doesn’t mean the project was all for naught, though, as the intention was never to make the next PGA champion, but to replicate the techniques for other robotic applications. “You can also transfer that to other problems, where you have some knowledge about the system and could model parts of it to obtain some data, but you can’t model everything,” Niklas Fittkau, the co-lead author of the paper, told IEEE Research.