Model Predictive Control with Variational Autoencoders for Signal Temporal Logic Specifications.

Area of Science:

• Control Systems Engineering
• Artificial Intelligence
• Robotics

Background:

• Dynamical systems often face complex control challenges with multiple, potentially conflicting rules.
• Existing control strategies struggle with prioritizing rules when simultaneous satisfaction is impossible.
• Human experts excel at managing such rule-based dilemmas, offering valuable insights for automation.

Purpose of the Study:

• To develop a novel control strategy synthesis method for dynamical systems with differential constraints.
• To address scenarios where not all specified rules can be simultaneously satisfied for task completion.
• To enable controllers to emulate human expert behavior in rule prioritization and management.

Main Methods:

• A learning-based Model Predictive Control (MPC) approach is proposed, integrating traditional control with machine learning.
• Rules are formally represented using Signal Temporal Logic (STL) formulas.
• A Conditional Variational Autoencoder (CVAE) learns a robustness margin from expert demonstrations to quantify rule satisfaction.

Main Results:

• The learned robustness margin guides the MPC process, facilitating rule prioritization and exclusion.
• The method successfully generates control behavior that mimics human experts in complex scenarios.
• Effectively manages rule-based dilemmas in simulated track driving environments.

Conclusions:

• The proposed learning-based MPC method offers a robust solution for control strategy synthesis in systems with conflicting rules.
• Emulating human expert decision-making enhances the adaptability and effectiveness of autonomous systems.
• This approach provides a framework for developing more intelligent and context-aware control systems.