Pathfinding in stochastic environments: learning vs planning.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Science

Background:

• Mobile robot navigation in complex environments faces challenges like partial observability and stochasticity.
• Dynamic environments with randomly appearing/disappearing obstacles are particularly difficult.
• Existing pathfinding methods struggle with unpredictable environmental changes.

Purpose of the Study:

• To propose and evaluate a stochastic formulation for mobile robot pathfinding in partially observable and dynamic environments.
• To compare planning-based and learning-based approaches for robust navigation.
• To provide insights into the scalability and performance of different navigation strategies.

Main Methods:

• Developed a stochastic pathfinding problem formulation considering dynamic obstacles and partial observability.
• Implemented a planning-based approach using a search-based planner with continuous path replanning.
• Implemented a learning-based approach utilizing recurrent neural networks for policy optimization.

Main Results:

• The learning-based approach demonstrated superior scalability with an increasing number of unpredictable obstacles.
• The planning-based approach proved more effective in environments with low stochasticity (near-deterministic conditions).
• Extensive empirical evaluation validated the performance trade-offs between the two methods.

Conclusions:

• Learning-based navigation is more robust and scalable for highly dynamic and uncertain environments.
• Planning-based navigation remains a viable option for more predictable and stable environments.
• The study offers practical guidance for selecting appropriate navigation strategies based on environmental characteristics.