Energy-Based Continuous Inverse Optimal Control.

Area of Science:

• Robotics
• Machine Learning
• Control Theory

Background:

• Continuous inverse optimal control aims to infer cost functions from expert demonstrations.
• Existing methods often struggle with the complexity of continuous control spaces and finite time horizons.

Purpose of the Study:

• To develop a novel framework for continuous inverse optimal control using energy-based models (EBMs).
• To enable learning of unknown cost functions from expert trajectories in a finite time horizon.

Main Methods:

• Utilizing an energy-based model (EBM) where expert trajectories are samples from a probability density.
• Implementing an "analysis by synthesis" scheme with Langevin dynamics and backpropagation through time for parameter learning.
• Exploring an optimization-based approximation and cooperative learning with a trajectory generator for enhanced efficiency.

Main Results:

• The proposed EBM framework successfully learns cost functions from expert demonstrations.
• Both the optimization approximation and cooperative learning strategies demonstrated improved efficiency.
• The methods were validated on autonomous driving tasks, showing effective cost function learning.

Conclusions:

• Energy-based models provide a robust framework for continuous inverse optimal control.
• The "analysis by synthesis" approach, enhanced by optimization or cooperative learning, offers an efficient solution.
• The learned cost functions are suitable for optimal control applications, particularly in autonomous driving.