Autonomous Navigation in Ice-Covered Waters with Learned Predictions on Ship-Ice Interactions

Ninghan Zhong, Alessandro Potenza, Stephen L. Smith

Autonomous Systems Laboratory

University of Waterloo

ICRA 2025

[arXiv]                            [GitHub]

Abstract

Autonomous navigation in ice-covered waters poses significant challenges due to the frequent lack of viable collision-free trajectories. When complete obstacle avoidance is infeasible, it becomes imperative for the navigation strategy to minimize collisions. Additionally, the dynamic nature of ice, which moves in response to ship maneuvers, complicates the path planning process. To address these challenges, we propose a novel deep learning model to estimate the coarse dynamics of ice movements triggered by ship actions through occupancy estimation. To ensure real-time applicability, we propose a novel approach that caches intermediate prediction results and seamlessly integrates the predictive model into a graph search planner. We evaluate the proposed planner both in simulation and in a physical testbed against existing approaches and show that our planner significantly reduces collisions with ice when compared to the state-of-the-art.

Method

We propose a deep learning framework that predicts the coarse dynamics of the ice motion in response to the ship actions through occupancy estimation. The learning process is guided with a novel physics-derived loss function tailored to our occupancy formulation.  To leverage the prediction results for planning, we present a simple yet empirically effective cost function based on occupancy maps to penalizes collisions. Further, we propose a graph search planner that seamlessly incorporates the learned model by caching intermediate predictions. 

Simulation Experiments

We conduct evaluations on an open-sourced autonomous ship-ice navigation simulator. The setup is consistent with the experimental platform at the National Research Council Canada ice tank in St. John’s, NL, featuring a 1:45 scale model vessel. Tests are conducted at 20%, 30%, 40%, and 50% concentrations, with 200 randomly generated environments per concentration. We evaluate our proposed planner against three baselines on the same set of trials, giving a total of 4 concentrations × 200 trials × 4 planners = 3200 experiments. These experiments demonstrate that our predictive planner, compared to the best-performing baseline: 

• Reduces ship kinetic energy loss due to collisions by up to 27%

• Reduces total ship-ice collision impulses by up to 19%

• Reduces approximated work required for navigation by up to 21%

• Maintains comparable travel distances

Autonomous Ship-Ice Navigation Dry-Land Testbed Experiment

We validate the proposed planner in a physical testbed that simulates autonomous ship-ice navigation at the University of Waterloo Autonomous System Laboratory.  The testbed features a 2.8 m × 1.6 m navigation environment. A TurtleBot3 Burger is used as a model vessel. We perform testbed evaluations in 20%, 30%, and 40% concentration environments. For each concentration, 40 trials are performed for each planner, giving a total of 3 concentrations × 40 trials × 4 planners = 480 experiments. Through these experiments, we show that our planner, compared to the best-performing baseline:

• Reduces approximated work required for navigation by up to 18%

• Maintains comparable travel distances