Overview
This paper introduces an observation-only autonomy framework designed for Unmanned Underwater Vehicles (UUVs) operating in dynamic underwater environments. The framework integrates persistent occupancy mapping, global clearance-aware planning, and risk-aware local control to enable robust UUV navigation under partial observability. The proposed pipeline leverages onboard sonar and depth image observations to construct occupancy maps. It adapts a clearance-constrained global planner (GP) to provide long-horizon structural guidance and incorporates a reinforcement learning (RL) policy for short-range tracking and reactive avoidance behaviors.
Approach
The system addresses decision-making under partial observability by learning a compact latent state representation. This representation is derived from onboard sensor data and encodes environmental structure, obstacle dynamics, and associated uncertainty. To enhance safety and training stability during the learning process, the framework utilizes Behavior Tree (BT) distillation, which includes staged supervision. A key feature is an uncertainty-calibrated distillation mechanism. This mechanism reweights teacher guidance dynamically using online latent-model uncertainty. The objective of this reweighting is to emphasize uncertain regimes during the learning process. The planning and local policy features explicitly retain time-to-collision (TTC) and clearance cues.
The methodology involved establishing a reproducible multi-seed evaluation protocol. This evaluation was conducted in a high-fidelity GPU-accelerated simulation environment, specifically using NVIDIA Isaac Sim. Performance of the proposed framework was benchmarked against two baseline approaches: a BT-only system and a standard RL system.
Findings
The evaluation results demonstrated improved robustness and safety of the proposed framework under dynamic conditions. The framework provides a general pipeline that combines a unified hybrid planning-learning architecture with a reproducible methodology. This combination supports robust UUV autonomy, particularly in scenarios characterized by partial observability.
Why This Matters
The presented framework offers a method for UUVs to navigate dynamic underwater environments with improved safety and robustness. By integrating diverse control and perception mechanisms, it addresses challenges associated with partial observability in underwater operations, potentially enhancing the reliability of autonomous underwater missions.