One important aspect in the field of legged robots is their robustness to unforeseen events during a mission. While walking on uneven ground, a legged walking robot is likely to fall. In order to continue its mission, the robot needs to provide a self-righting behavior, which succeeds from any initial position.
The four legged robot Charlie is biologically inspired by an ape. Its features are different four legged walking gaits, an artificial spine and different proprioceptive and interoceptive sensors.
This master thesis targets to develop a self-righting behavior for Charlie. As Charlie has 39 active DOF in total, the resulting state space is large. To approach the task, predefined robot states, as well as parameterized actions can be used. Also state- and action space reduction and hierarchical reinforcement learning methods are promising to learn a self-righting procedure. It will be investigated, if Charlie's artificial spine can improve the self-righting behavior. At first, the MARS simulator together with a model of Charlie will be used to develop the core behavior.  When the self-righting behavior succeeds in simulation, it will be transferred to a ROCK-Library and tested in the MARS simulator as well as on the real system.