Constrained stochastic optimal control with learned importance sampling: A path integral approach.

Area of Science:

• Robotics
• Control Theory
• Artificial Intelligence

Background:

• Modern robots require robust operation in unpredictable environments.
• Existing control methods struggle with high-dimensional systems and partial uncertainty.
• Real-time control in unknown scenarios remains a significant challenge.

Purpose of the Study:

• To propose a novel algorithm for controlling high-dimensional robotic systems in partially unknown environments.
• To extend stochastic optimal control with constraint-handling capabilities for enhanced robustness.
• To enable real-time robotic control applications using sampling-based methods.

Main Methods:

• Utilizes the path integral formulation of stochastic optimal control.
• Infers optimal control inputs from stochastic system dynamics rollouts simulated via a physics engine.
• Employs importance sampling and constraint handling to reduce sampling complexity for online control.

Main Results:

• Demonstrates the feasibility of real-time control applications through effective sampling complexity reduction.
• Shows that the proposed algorithm provides an additional layer of safety and robustness, even surpassing ancillary controllers.
• Validates the algorithm's effectiveness on diverse robotic systems, including hardware experiments on a quadrupedal robot.

Conclusions:

• The developed stochastic control algorithm offers a robust solution for controlling complex robotic systems in challenging environments.
• The method facilitates both online refinement and offline learning of control policies, enhancing adaptability.
• This approach significantly advances the capabilities of robotic systems operating under uncertainty.