The human machine interface (HMI) refers to a paradigm in which the users interact with external devices through an interface that mediates the information exchanges between them and the device. In this work we focused on a HMI that exploits signals derived from the body to control the machine: the body machine interface (BMI). It is reasonable to assume that signals derived from body movements, electromyography activity, as well as brain activity, have a non-linear nature. This implies that linear algorithms cannot exploit all the information contained in these signals. In this work we proposed a new BMI that maps electromyographic signals into the control of a computer cursor by using a new non-linear dimensionality reduction algorithm based on autoassociative neural network. We tested the system on a group of ten healthy subjects that, controlling this cursor, performed a reaching task. We compared the result with the performance of an age and gender matched group of healthy subjects that solved the same task using a BMI based on a linear mapping. The analysis of the performance indices showed a substantial difference between the two groups. In particular, the performance of the people using the non-linear mapping were better in terms of time, accuracy and smoothness of the cursor's movement. This study opened the way to the exploitation of non-linear dimensionality reduction algorithms to pursue a new and effective clinical approach for body-machine interfaces.