Formation Control With Collision Avoidance Through Deep Reinforcement Learning Using Model-Guided Demonstration.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Formation control with collision avoidance (FCCA) in dynamic environments is challenging.
• Existing methods often address formation maintenance and collision avoidance separately.
• Follower robots must simultaneously maintain formation and avoid collisions.

Purpose of the Study:

• To propose a novel deep reinforcement learning (RL) based method for FCCA.
• To develop a two-stage training framework combining imitation learning (IL) and RL.
• To enhance the perception and generalization capabilities of the control system.

Main Methods:

• A two-stage training framework: imitation learning (IL) followed by reinforcement learning (RL).
• IL stage utilizes a model-guided approach with consensus theory and optimal reciprocal collision avoidance.
• RL stage employs a compound reward function and a formation-oriented network with Long Short-Term Memory (LSTM).
• Transfer training approach is used for improved generalization.

Main Results:

• The proposed method successfully generates collision-free, time-efficient paths.
• The formation-oriented network effectively perceives uncertain numbers of obstacles.
• The approach demonstrated strong generalization across different scenarios.
• Effectiveness and practicality validated through extensive simulations and experiments.

Conclusions:

• The deep reinforcement learning method offers an effective solution for FCCA problems.
• The two-stage training and specialized network architecture improve performance and generalization.
• The approach is practical and validated on a real-world multi-omnidirectional-wheeled car system.