Policy-Guided Model Predictive Path Integral for Safe Manipulator Trajectory Planning

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Safe trajectory planning for manipulators in complex environments faces challenges with hard-constraint enforcement and environmental generalization.
• Existing methods struggle to balance real-time optimization with robust safety guarantees.

Purpose of the Study:

• To propose a novel Policy-Guided Model Predictive Path Integral (PG-MPPI) planning framework for manipulator trajectory planning.
• To improve hard-constraint enforcement and environmental generalization capabilities for safe robot operation.

Main Methods:

• Developed a Constraint-Discounted Soft Actor-Critic (CD-SAC) offline learning policy incorporating a configuration-space distance field for obstacle avoidance.
• Integrated the offline policy with Model Predictive Path Integral (MPPI) for guided online sampling and optimization.
• Implemented a Control Barrier Function (CBF) safety filter for real-time control command revision and constraint satisfaction.

Main Results:

• The PG-MPPI algorithm demonstrated superior performance in collision-free target reaching success rates compared to other algorithms.
• The proposed method ensures trajectory smoothness and feasibility in multi-obstacle scenarios.
• The framework exhibits strong adaptive capacity to complex environments with unknown obstacle configurations.

Conclusions:

• The PG-MPPI framework offers an efficient and safe solution for autonomous manipulator operation in complex, dynamic environments.
• The integration of RL and MPC with CBF significantly enhances safety and generalization.
• This approach advances the state-of-the-art in safe robot motion planning.