Imitation Learning

Johannes Busch

Machine Learning for Robotics - SuSe 24

Learning, Adaptive Systems, and Robotics (LASR) Lab

Practice 12:

v1.0

Imitation Learning

• In Imitation Learning we learn policies from expert data instead of generating data from interaction (RL).
• Expert data can be expensive to obtain.d

Behavior Cloning

• Behavior Cloning uses SL to learn a policy: action = f(state) from a dataset of demonstrations (state, action).
• For this we have to rely on high-quality data.

Offline Reinforcement Learning

• Offline RL combines ideas from RL and IL.
• It uses datasets of sub-optimal demonstrations (state, action, reward) to extract high-performing policies.

1.0: 

• Offline RL combines ideas from RL and IL.
• It uses datasets of sub-optimal demonstrations (state, action, reward) to extract high-performing policies.