ETD

• deep learning
• deep reinforcement learning
• elastic
• evolutionary strategy
• explosive motion
• jump
• learning
• locomotion
• quadruped
• reinforcement learning
• soft
• springs

The aim of this thesis is to use Deep Reinforcement Learning for training a bio-inspired quadruped to acquire explosive motion skills that could be used for navigation in challenging natural environments.
The jumping task has been taken into account, either jumping in place or jumping forward.

The training method proposed combines evolutionary strategy and deep reinforcement learning algorithms in two steps. During the first one, ARS (Augmented Random Search) is used in combination with a sparse reward and deterministic behavior. In the second one, an imitation learning simplified approach is used to train a more complex neural network with PPO (Proximal Policy Optimization) that is successively retrained with a task-related reward function.

In the end, the agent learns how to jump and softly land starting and ending in a default position exploiting the compliant elements (joint-level parallel springs) for better performances.