• Anymal
• Autonomous Navigation
• Behavioral Cloning
• Decision Transformer (DT)
• Experimental Validation
• Feed-Forward Network
• Hardware Constraints
• Imitation Learning
• Long Short-Term Memory (LSTM)
• Multimodal Perception
• Navigation Dataset
• Quadrupedal Robotics
• Teleoperation

This thesis addresses the autonomous navigation of a quadruped robot through Imitation Learning, with the goal of translating human intuition into autonomous action. The end-to-end approach was validated by comparing three different neural network architectures. The methodology was based on the creation of a large-scale dataset, collected from 36 users via an immersive teleoperation interface that combines camera and LiDAR data. Following a preprocessing pipeline, this data was used to train, via Behavioral Cloning, a reactive network (FFN) and two memory-based architectures (LSTM, Decision Transformer), using a perception front-end with pre-trained models and a custom CNN for feature compression. The methodology was experimentally validated on the quadrupedal robot ANYmal, showing that memory-based models overcome the limitations of reactive ones, and posing the groundwork for autonomous end-to-end navigation.