In most research into learning on autonomous robots, only high level tasks like navigation are learned with more or less task specific learning systems. In this paper research is conducted into learning very basic, low level tasks like forward movement, backward movement, halting behaviour and phototaxis. A learning system based on very simple parallel processes and some ideas from genetic algorithms is implemented on a real robot and it is shown that certain behaviours can be learnt extremely quickly. This confirms results from simulations of the same system. For other real world tasks the learning system is shown to be too simple, but an extension is proposed that should overcome these problems.