Online Navigation Planning for Long-term Autonomous Operation of Underwater Gliders
Authors: Victor-Alexandru Darvariu, Charlotte Z. Reed, Jan Stratmann, Bruno Lacerda, Benjamin Allsup, Stephen Woodward, Elizabeth Siddle, Trishna Saeharaseelan, Owain Jones, Dan Jones, Tobias Ferreira, Chloe Baker, Kevin Chaplin, James Kirk, Ashley Iceton-Morris, Ryan D. Patmore, Jeff Polton, Charlotte Williams, Christopher D. J. Auckland, Rob A. Hall, Alexandra Kokkinaki, Alvaro Lorenzo Lopez, Justin J. H. Buck, Nick Hawes
Organizations: Oxford Robotics Institute, Department of Engineering Science, University of Oxford, Oxford, UK · Optiver, Amsterdam, The Netherlands · Stateful Robotics, Oxford, UK · National Oceanography Centre, Southampton, UK · School of Environmental Sciences, University of East Anglia, Norwich, UK · Scottish Association for Marine Science, Oban, UK
Underwater glider robots have become indispensable for ocean sampling, yet fully autonomous long-term operation remains rare in practice. Although stakeholders are calling for tools to manage increasingly large fleets of gliders, existing methods have seen limited adoption due to their inability to account for environmental uncertainty and operational constraints. In this work, we demonstrate that uncertainty-aware online navigation planning can be deployed in real-world glider missions at scale. We formulate the problem as a stochastic shortest-path Markov Decision Process and propose a sample-based online planner based on Monte Carlo Tree Search. Samples are generated by a physics-informed simulator calibrated on real-world glider data that captures uncertain execution of controls and ocean current forecasts while remaining computationally tractable. Our methodology is integrated into an autonomous system for Slocum gliders that performs closed-loop replanning at each surfacing. The system was validated in two North Sea deployments totalling approximately 3 months and 1000 km, representing the longest fully autonomous glider campaigns in the literature to date. Results demonstrate improvements of up to 9.88% in dive duration and 16.51% in path length compared to standard straight-to-goal navigation, including a statistically significant path length reduction of 9.55% in a field deployment.
Figures & tables
Figure 1: Slocum glider in operation during the eSWEETS 3 field deployment. The proposed Glider Autonomy Long-term Planning Engine system was used to pilot gliders autonomously in two long-term deployments totalling approximately 3 months and 1000 km.
Figure 2: An illustration of basic glider operation and the relevant control parameters for Slocum gliders. These include the yo floor zbottom , ceiling ztop , climb pitch ϑclimb , dive pitch ϑdive , buoyancy change Δbuoy , and number of yos nyos ( 3 yos are completed in this dive).
Figure 3: The online planning loop of the GALE system and key considerations. Upload Science Data and Plan Next Dive occur simultaneously while the glider is at the surface.
Figure 4: An illustration of the key inputs and outputs of the glider dive simulator. The glider icon represents the dive initial location , the arrow indicates its direction , and background arrows correspond to ocean current forecasts over several depth levels. The grey circles are the outputs of the simulator indicating possible surfacing locations , i.e., stochastic MDP transitions. The red triangle is the true surfacing location .
Figure 5: An illustration of a search tree built by the planner. Circles represent state nodes and squares represent action nodes. Action nodes can have mulitple children as MDP transitions T are stochastic. Q is the expected cost of an action. At the root, diving at −30 degrees relative to the goal is estimated to lead to the lowest expected cost. The search tree is several levels deep and explored asymmetrically by the planner to balance exploration and exploitation. Starred nodes denote terminal states in which the glider reaches the goal.
Figure 6: Illustration of the proposed action space formulation (left), which uses bearings α that are relative to the goal. This leads to a lower branching factor than deciding headings ψ directly (right) while retaining relevant actions.
Figure 7: Visualisation of waypoint list generation. The first waypoint corresponds to the planner action (here, α=−30 ). The first backup waypoint is set to be the current goal as otherwise the goal would be overshot. The second backup waypoint is aimed straight at the next goal.
Figure 8: Diagram illustrating the components and steps of the two system workflows. The Mission Configuration Workflow (right) is executed once at the start the mission, while the Online Planning Workflow (left) is executed every few hours upon the surfacing of the glider.
Deployment
Goal 1
Goal 2
Dist.
ρ
ρwpt
nbck
MOGli
59.32∘ N 0.35∘ W
59.32∘ N 0.65∘ W
17.0
2.0
7.0
2
eSWEETS 3 N
57.089∘ N 1.856∘ W
57.062∘ N 1.698∘ W
10.0
1.0
7.0
2
eSWEETS 3 S
56.929∘ N 1.945∘ W
56.903∘ N 1.787∘ W
10.0
1.0
7.0
2
Table 1: Deployment goal locations and parameters. All distances are given in km.
Deployment
Transects completed
Duration (days)
Total dives
Distance travelled (km)
MOGli
41
42
343
665
eSWEETS 3 North
20
27
150
173
eSWEETS 3 South
23
22
147
207
Table 2: Deployment summary statistics.
Figure 9: Field deployment locations in the North Sea.
Figure 10: Learning curve showing the best value of J(Θ) found as a function of Bayesian optimisation iterations using the MOGli training data.
Metric
MOGli
eSWEETS 3 North
eSWEETS 3 South
True dive distance (m)
2733
5650
4679
Simulated dive distance (m)
2119
4991
4362
Distance delta (m)
1448
1555
1080
True dive duration (s)
10421
15923
12935
Duration delta (s)
345
1481
1377
Mean log-likelihood ℓˉ
-9.29
-78.67
-133.90
Table 3: Comparison of simulated and true dive metrics averaged over all dives in the real-world deployments.
Figure 11: Progress of the proposed simulator parameter optimisation procedure (Section 4.D ) on one of the dives in the MOGli deployment. Simulator parameters are progressively adjusted such that the surfacing locations generated by the simulator (grey circles) best match the true surfacing location (red triangle).
Deployment
Scenario
Algorithm
Mean Duration (hrs)
Mean Number Dives
Mean Path Length (km)
MOGli
Fav.
STG
22.48 ± 0.30
7.68 ± 0.10
17.31 ± 0.05
Planner
22.13 ± 0.30
7.57 ± 0.10
16.79 ± 0.06
Neutral
STG
22.93 ± 0.31
7.87 ± 0.11
17.66 ± 0.05
Planner
22.38 ± 0.30
7.68 ± 0.10
16.87 ± 0.06
Unfav.
STG
23.60 ± 0.34
8.10 ± 0.12
17.85 ± 0.06
Planner
22.84 ± 0.32
7.80 ± 0.11
16.92 ± 0.06
Table 4: Results comparing the Straight-to-Goal and Planner algorithms across several scenarios in simulation. Bold indicates the better-performing algorithm.
Figure 12: Transects completed by the proposed GALE system in the MOGli field deployment using the Planner and the Straight-to-Goal baseline. Only transects in which all transfers of control instructions succeeded are shown. The goal radii are represented by two dotted circles around the goals. The Planner results in transects with shorter minimum, mean, and maximum durations. Transects produced by the two methods are also discernible visually. The transects completed by the Planner have statistically significantly shorter path lengths and avoid overshooting the western goal.
Deployment
Algorithm
Mean Duration (hrs)
Min Duration (hrs)
Max Duration (hrs)
Mean Number Dives
Min Number Dives
Max Number Dives
Mean Path Length (km)
MOGli
Straight-to-Goal
23.09
16.00
31.70
8.45
6
11
22.58
Planner
22.68
14.35
26.98
8.28
5
10
20.52
eSWEETS 3 North
Straight-to-Goal
40.38
19.80
83.54
9.43
5
18
49.10
Planner
28.88
18.05
43.10
6.46
4
10
38.36
eSWEETS 3 South
Straight-to-Goal
21.37
10.75
54.64
6.00
3
14
29.29
Planner
26.82
10.44
47.55
7.29
3
13
32.95
Table 5: Results comparing the Straight-to-Goal and Planner algorithms in the field deployments. Bold denotes the better-performing algorithm for each deployment and metric.
Figure 13: Proportions of actions chosen by the Planner during the three field deployments, split between the East and West directions of travel. The stronger currents of the eSWEETS 3 region resulted in a wider distribution of action frequencies.
Autonomous underwater navigation develops methods for underwater vehicles to localize, plan routes, and move safely despite uncertain sensing, currents, and communication.
Centre for Systems and Control, Indian Institute of Technology Bombay, Mumbai · Institute for Control, Optimization, and Networks, Purdue University, West Lafayette, USA