Operating unmanned ground vehicles (UGVs) in partially observable battlefields requires balancing movement efficiency and proxy risk exposure. Traditional shortest-path algorithms typically struggle here, as their use of purely geometric distances often yields tactically vulnerable routes. We confront this challenge by introducing a risk-aware path-planning framework driven by Trajectory-Value Factorized Recurrent Proximal Policy Optimization (TVF-RPPO). Our integrated navigation pipeline combines a SwiftFormer perception unit to map terrain-specific costs and risks with a GRU-based belief state that continuously tracks observation history. Unlike standard approaches that fuse competing goals into a single scalar reward, TVF-RPPO explicitly splits value estimation into separate time-efficiency and risk-avoidance channels. This structural shift is the key mechanism that enables the network to efficiently regulate speed and safety. During our 2D evaluations involving static, dynamic, and active threat scenarios, the technique reduced proxy risk exposure compared with baseline planners while maintaining competitive mission-completion performance in the tested simulation scenarios. A major practical benefit of this factorization is its adjustability. By simply tuning a single risk-weight parameter during inference, the UGV can immediately switch between cautious and aggressive maneuvers without additional training.
Kim et al. (Tue,) studied this question.