However, there are three aspects to the problem, two of which are extremely difficult!

1. Patrolling the protected space: a lot harder than it looks, but by now this is basically a solved problem using off-the-shelf algorithms.

2. Dealing with the weed. This is much harder than it looks. Quite a few weeds can be uprooted by firmly grasping the base near the ground and steadily pulling upwards. However, there are many others that require particular actions, and in some cases these actions are very complicated for a robot. For example, around here we have a particularly nasty weed called “cat-head.” The basic action is as above, but first you have to disentangle all the stolons form the surrounding plants. Failing to do so before uprooting the main body will cause stolons to tear off and scatter viciously thorned seeds into the crop.

3. The hardest problem, though, is reliably and accurately distinguishing weeds from wanted plants. There doesn’t seem to be any way to do this except a very difficult image recognition problem: a problem where all the general features are similar or identical between the “treat” and “no treat” cases, and only the overall pattern counts. This is a notoriously difficult problem, similar to facial recognition, and equal error rates are on the order of 5 ~ 10%. If every pass of the weeding robot resulted in 5% of weeds missed and 5% of crop destroyed, the customer would be very, very unhappy!

For example, one robot may trim, and another haul, and a third lift.

I would like a wheelchair like that when i’m 80.

Are some of those Designs from the FH Johaneum in Graz/Austria?