Description

This thesis explores the development of a hybrid training pipeline for quadrupedal locomotion on the Solo12 robot, combining both simulation-based learning and real-world experimentation. The objective is to leverage the scalability and safety of simulation together with the physical fidelity and grounding offered by hardware trials. The work integrates modern reinforcement learning methods within a unified framework that enables policies to be trained, validated, and progressively deployed on the real robot.

The work is under the scope of the following projects:

• SGR RAIG: Mobile Robotics and Artificial Intelligence Group (web)
• LENA: Lifelong navigation learning using human-robot interaction (web)
• TORNADO: foundaTion mOdels for Robots that haNdle smAll, soft and Deformable Objects (web)