Mobile autonomous robots are increasingly conquering many fields of applications.
 The key feature of those systems is their autonomy due to their robustness against unpredictable changes of the environment or the mission goal, eliminating the need of intervention of a human operator.


Legged robots are commonly seen as the most adaptable systems, as legs enable a larger variety of movements than for example wheels. 
One important ability of legged robots for increasing their robustness, is the recovery from falling, by providing a recovery strategy, for self-righting and standing up.
In this work, the applicability of machine learning methods to the challenge of finding a self-righting behavior for a representative of this type of robots, the kinematically complex, biologically inspired, four-legged robot Charlie is investigated.
 Of special interest is hereby the influence of an artificial spine joint on such a behavior.
At first, a suitable behavior representation is developed, to allow the robot to transition from a supine lying pose to a prone lying pose, and finally to its four-legged standing pose.

A fitness function is developed with regard to the objective of self-righting, taking different quantities from the simulation into account, such as orientation, collision forces, and actuator current, enabling to evaluate behaviors. The first handcrafted approach, is developed as reference and proof of concept behavior by utilizing human expert knowledge on standing up movements and intuition. The second, optimized approach, combines the first, reorienting part of the previously created trajectories with the optimization algorithms cmaes and pso, resulting in optimized movements. The third approach, referring to as evolved, is realized without prior knowledge. Optimization algorithms are applied to the evolved approach as well, targeting to find new reorienting trajectories. In order to prove that the found behavior representation presents a valid solution to the problem, both the real and simulated versions of the system are used during the development. The results clearly suggest, that machine learning methods are a promising approach to optimize self-righting behaviors. They also show, that the application of such methods is not trivial, due to large state- and action spaces.