Introduction
Autonomous navigation in dynamic environments typically relies on a continuous stream of high-fidelity sensor data. However, in many real-world scenarios—such as deep-ocean exploration—Unmanned Underwater Vehicles (UUVs) only receive intermittent updates about nearby obstacles when they surface to communicate. Once submerged, these vehicles must navigate using "stale" data that may be several minutes old. This research introduces a hybrid planning framework designed to maintain safety under these conditions by approximating computationally expensive Reachability Analysis (RA) through a compositional neural network (NN) controller.
The Challenge: Uncertainty and Computation
Navigating with intermittent information presents two primary hurdles:
- Uncertainty Management: Because obstacle data is delayed, the robot cannot know an obstacle's exact position. Instead, it must account for the "forward reachable set"—the entire spatial envelope an obstacle could occupy given its last known state and possible dynamics.
- Computational Constraints: While Reachability Analysis provides rigorous safety guarantees, calculating these sets in real-time is mathematically intensive. Traditional tools are often too slow to function within a robot's high-frequency control loop.
Methodology
The authors propose a framework that shifts the computational burden of safety verification to an offline training phase, enabling real-time execution.
Reachability and Potential Fields
Offline, the system uses tools like Flow* to calculate the forward reachable sets of obstacles, accounting for process noise. These sets are then integrated into a Potential Field (PF) framework. In this model, the reachable sets act as sources of a "repulsive force" that pushes the robot away from potential collision zones, while an "attractive force" pulls the robot toward its goal.
Neural Network Approximation
To achieve real-time performance, a neural network is trained to learn the gradient of this repulsive field. The NN takes the obstacle's last known state and the time elapsed since the update as inputs, outputting a reactive avoidance vector. By including "time since last update" as a primary feature, the controller naturally becomes more conservative as the data ages and the reachable sets expand.
Compositionality
A key innovation of this approach is its compositionality. Because the gradients of potential fields are additive, the NN is trained only on a single-obstacle scenario. At runtime, the robot simply sums the avoidance vectors for all detected obstacles. This allows the system to scale to complex environments with numerous obstacles without requiring additional training.
Experimental Results
The framework was validated through high-fidelity ROS Gazebo simulations for UUVs and physical experiments using Clearpath Jackal and TurtleBot UGVs.
- Computational Efficiency: The NN controller achieved an average runtime of 27ms per iteration. This is a significant improvement over traditional Reachability Analysis (using Flow), which averaged 67 seconds*—a speedup of approximately 2,500 times.
- Safety Performance: In simulations of a UUV crossing a shipping channel with 9-minute-old data, the NN planner maintained a minimum safety distance of 7.5m.
- Comparative Advantage: The proposed system significantly outperformed "Noiseless Planners" that ignore uncertainty. While the noiseless models frequently resulted in collisions by planning based on outdated positions, the NN controller successfully accounted for potential obstacle deviations.
Limitations and Future Directions
While the framework provides a robust solution for intermittent data, it currently assumes that the robot has general knowledge of the obstacles' control policies (e.g., maintaining a constant heading). If an obstacle behaves erratically, safety guarantees may be compromised. Additionally, safety is only guaranteed up to a specific time threshold $T$, beyond which the reachable sets may become too large for effective navigation.
Future work will focus on incorporating the robot’s own localization uncertainty and developing "confidence monitoring" systems to help autonomous vehicles determine exactly when they must seek fresh sensor updates to ensure continued safe operation.
Summary updated