Representation Learning and Reinforcement Learning for Dynamic Complex Motion Planning System.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Indoor motion planning is challenging due to dense, unpredictable moving obstacles, hindering classical algorithms.
• Reinforcement learning (RL) offers safe multiagent robotic motion planning but suffers from slow convergence and suboptimal results.
• Existing RL methods struggle with efficient learning in dynamic, cluttered environments.

Purpose of the Study:

• To develop a more efficient and effective hybrid motion planning algorithm for indoor environments.
• To address the convergence challenges of current reinforcement learning algorithms in multiagent robotic systems.
• To improve the safety and speed of robotic navigation in the presence of dynamic obstacles.

Main Methods:

• Implemented a discrete soft actor-critic (SAC) algorithm for discrete action spaces.
• Optimized data quality using attention-based long short-term memory (LSTM) encoding.
• Introduced a novel data replay method combining online and offline learning.

Main Results:

• The proposed ALN-DSAC algorithm demonstrates superior convergence compared to state-of-the-art methods.
• Achieved nearly 100% success rate in motion planning tasks.
• Reduced the time required for robots to reach their goals.

Conclusions:

• ALN-DSAC offers a significant advancement in robotic motion planning, particularly for dynamic indoor environments.
• The hybrid approach effectively combines RL and representation learning for improved performance.
• This algorithm provides a safer and more efficient solution for complex navigation tasks.