Risk-aware tactical path planning in partially observable environments via trajectory-value factorized recurrent PPO.

Area of Science:

• Robotics
• Artificial Intelligence
• Path Planning

Background:

• Operating unmanned ground vehicles (UGVs) in partially observable battlefields presents a challenge in balancing movement efficiency with risk exposure.
• Traditional shortest-path algorithms often produce tactically vulnerable routes due to reliance on purely geometric distances.

Purpose of the Study:

• To develop a risk-aware path-planning framework for UGVs in uncertain battlefield environments.
• To improve navigation by explicitly considering both movement efficiency and proxy risk exposure.

Main Methods:

• Introduction of Trajectory-Value Factorized Recurrent Proximal Policy Optimization (TVF-RPPO) for risk-aware path planning.
• Integration of a SwiftFormer perception unit for terrain cost and risk mapping.
• Utilization of a GRU-based belief state to track observation history.
• Explicitly splitting value estimation into time-efficiency and risk-avoidance channels within TVF-RPPO.

Main Results:

• The TVF-RPPO framework reduced proxy risk exposure compared to baseline planners in 2D evaluations.
• The system maintained competitive mission-completion performance across static, dynamic, and active threat scenarios.
• The approach demonstrated effective regulation of speed and safety through its dual-channel value estimation.

Conclusions:

• The proposed TVF-RPPO framework offers an effective solution for risk-aware path planning in challenging environments.
• The explicit factorization of value estimation allows for adjustable risk-taking behavior without retraining.
• This method provides a practical benefit through a single risk-weight parameter for switching between cautious and aggressive maneuvers.