This is the project from the course "Advanced Deep Learning in Robotics"
In this project, we basically reimplemented the algorithm from paper "Meta Reinforcement Learning for Sim-to-real Domain Adaptation".
To be more specific, we :
• Wrote PPO algorithm from scratch
• Wrote Reptile algorithm from scratch
• Wrote Pseudo MAML algorithm proposed in the literature from scratch
• Modify some classical Pybullet-Gym environments to conduct a series of experiments for model evaluation
To train your agent using pseudo MAML/PPO, please check the code in main.py
To train your agent using Reptile, please check the code in reptile_rl.py
Halfcheetah Environments:
Randomized Parameters:
*Joints Coef: offset 30%
*Dynamics Coef: offset 30%
DoubleInvertedPendulum Environments:
Randomized Parameters:
*Gravity: 1 -- 20
*Torque Factor: 50 -- 500
