CONTENTS

Presentation

[Slides]

ANSANGHO _ FAULT-TOLERANT CONTROLOFA QUADRUPED ROBOT.pdf

https://youtu.be/eeZr2bM1Hv0

What is DDPG? (Deep Deterministic Policy Gradient)

1. OverviewDeep Deterministic Policy Gradient (DDPG) is a model-free, off-policy reinforcement learning algorithm based on the Actor-Critic architecture. It combines the success of Deep Q-Networks (DQN) with the ability to handle continuous action spaces, making it a standard choice for high-dimensional robotic control tasks.

2. Why DDPG for Quadruped Robots? The primary reason for selecting DDPG for this project is its capability to handle Continuous Action Spaces.

• Discrete vs. Continuous: Traditional algorithms like DQN work well for discrete actions (e.g., specific buttons on a controller: Up, Down, Left, Right). However, a quadruped robot requires precise, continuous control of joint torques (e.g., applying exactly 5.2 Nm of torque to a knee joint).
• Precision Control: DDPG can output specific continuous values for the actuator inputs, allowing for the smooth and complex coordinated movements required for stable walking.

3. Key Architecture: Actor-Critic DDPG employs two neural networks that work in tandem: