• deep reinforcement learning
• learning
• motion generation
• robot
• training algorithms

This thesis presents a method for generating motion references for a generic robot providing it with velocity commands. The aim is to generalize Deep Reinforcement Learning techniques for mobility tasks, regardless of the robot type, environment, or specific objective. The methods will be tested on a quadrupedal robot with eight degrees of freedom. The thesis develops training and validation methodologies for neural networks capable of walking in unstructured environments with obstacles, changes in physical properties and disturbances. The work is validated through a Validation Pipeline ad hoc developed for mobility tasks. The thesis exposes the intermediate and final results obtained with various policies to highlight the common inconveniences found during a policy's synthesis and can serve as a guideline, in the Appendix, a recap table with the failure type, solutions, and references to the chapters where the solutions are better explained can be found.