This repo serves as a compilation of information, links, and code that is relevant for the use and reproduction of the datasets gathered by Niklas Rolleberg, Clemens Deutsch, and Aldo Terán Espinoza (from the Centre for Naval Architecture at the KTH Royal Institute of Technology in Sweden) using SMaRC's AUV LoLo at the Baltic Sea Center's Askölaboratoriet.
The datasets contain both the raw measurements and ground truth trajectories from our agents during a series of dynamic rendezvous (see below for a thorough description of the scenario). The ground truth in the datasets was computed by following the methods and procedure published in our paper:
@INPROCEEDINGS{11104553,
author={Terán Espinoza, Aldo and Terán Espinoza, Antonio and Deutsch, Clemens and Rolleberg, Niklas and Folkesson, John and Sigray, Peter and Kuttenkeuler, Jakob},
booktitle={OCEANS 2025 Brest},
title={A Consistent Dataset for Dynamic Underwater Proximity Operations},
year={2025},
volume={},
number={},
pages={01-09},
keywords={Navigation;Sea measurements;Optical variables measurement;Position measurement;Probabilistic logic;Trajectory;Vehicle dynamics;State estimation;Robots;Optimization;underwater docking;autonomous underwater vehicles;relative navigation;proximity operations;state estimation},
doi={10.1109/OCEANS58557.2025.11104553}}
If you use these datasets in your research, please use the reference above.
The big picture of our research is to enable autonomous proximity operations using AUVs. The scenario of an arbitrary prox-op consists of two agents, a chaser agent (an AUV in our case), and a target agent that can be anything from a static docking station, to a dynamically active mothership; the chaser navigates through a set of phases with specific sensor modalities and challenges before reaching its target. This framework is borrowed from spacecraft robotic prox-ops, and it’s used to organize and tackle the different problems that arise during the different phases in a holistic manner. The figure below depicts the different phases of a proximity operation.
The datasets presented here were gathered with the main purpose of using them to verify state estimation and target tracking algorithms that solve the inference problems that arise when executing proximity operations with a collaborative target underwater.
The experiments were carried out just outside of the Askolaboratoriet research station, in a bay that is protected from major currents and wind.
The objective of the experiments was to simulate an underwater scenario where a mothership (target) rendezvous with an AUV (chaser), with the goal of, for example, transferring data through an optical communication link without the need to stop or hard-dock. Thus, throughout the experiments, both vehicles are (almost) always in motion.
As a surrogate for the mothership, we instrumented our service boat with (1) an SBG Ellipse-D AHRS for navigation data, (2) an Evologics SR17 USBL positioning system to track and communicate with the AUV during the long-distance phase, and (3) an array of BlueRobotics Lumen underwater lights that serve as a fiducial marker during the terminal phase of the operation (see image below). This sensor suite was interfaced to a laptop using ROS onboard the service boat.
The AUV LoLo is equipped with a plethora of scientific payloads, however, for the sake of our experiments we only gathered the data from the sensor suite used for relative navigation, composed of: (1) an iXBlue (now Exail) ROVINS inertial navigation system (INS) with a Nortek DVL, (2) an Evologics SR17 transponder and acousitc communication modem, and (3) a Flir Blackfly S GigE - BFS-PGE-27S5C-C machine vision camera with a Kowa LM5JCM | 2/3" 5mm wide angle lens inside a BlueRobotics 2" pressure vessel with an optically clear dome lid. LoLo has an onboard "scientist" computer, a Jetson AGX Orin, that is in charge of data acquisition also using ROS.
Important
Because the AUV is NOT equipped with USBL positioning (only the unit on the boat is capable of range-bearing measurements of the transponder on LoLo), the measurement obtained by the service boat's USBL unit are sent to the AUV via acoustic link. Thus, the USBL fixes that the AUV acquires of the service boat are bound to have a (probably small but unknown) delay.
Three deployments were carried out using the experimental setup above (it was like 10 in reality but the stars aligned only in 3 of them). The deployments consisted on LoLo diving at an (almost) constant depth and velocity in a ca. 500m circuit (see figure below). The datasets differ in the number of laps that LoLo did during the deployment.
Since LoLo was running "open-loop," that is, without having any feedback of our whereabouts in the service boat, it was up to us to use the USBL positioning system to find and get close enough for the AUV to capture images of the light fiducials--effectively generating data of a simuluated long- to terminal-phase proximity operation. We carried out the rendezvous maneuver in a series of different manners: (1) following at acoustic distance and then closing up in a straight line, (2) following parallel to the AUV and then merging to present the lights to the camera, (3) intercept from a long distance and in an angle, and (4) combinations of the above.
To compute the so-called ground truth in our datasets, we used a factor graph optimization method to jointly estimate the most consistent trajectories for both agents using all the available global and relative navigation measurements. These methods are described in detail in our publication cited above.
The data consists of three different datasets composed of the merged rosbags of both vehicles. Because both systems were running on network time, always having internet connection (except when underwater of course), we assumed their clocks to be synced enough for a simple merge of the timestamped data to be accurate. Careful qualitative assesment of the post-processed and merged data seems to confirm that there are no big, perceivable, delays.
The topics in the rosbags are divided into a series of namespaces, for LoLo's data:
/lolo/core/: raw data mainly from the navigation sensors./lolo/dr/: processed navigation data (dr = dead reckoning), essentially navigation data in friendlier rosmsgs./lolo/sensors/: payload sensor information; the most important here is thecamera/image_monotopic with the compressed unprocessed image of the backwards-facing camera./lolo/perception/: "front-end" post-processed data from the acoustic comms (USBL fix), and the calculated relative pose of the light fiducials wrt the AUV's camera.
The service boat's data follows:
/evologics/: raw data from the Evologics USBL positioning system./sbg/: data from the SBG AHRS with their own message types./imu/: data from the SBG AHRS in standard rosmsgs./service_boat/: processed navigation and USBL data.
Topics with prefix /gt are the computed ground truth trajectories for LoLo and the service boat.
You have to make sure to have the following dependencies installed in your workspace:
- The dmac ROS driver: we need the message definitions for the Evologics USBL.
- Same for the AHRS data, you need the sbg ROS driver.
To set up the TF tree and the robot model LoLo, you need
and for the service boat
- the asko-2024 branch of the usbl_tools package.
Download the datasets from our OneDrive server here.
Note
We have issues hosting our datasets on the KTH OneDrive service, if the link above does not work, feel free to open an issue and we'll address it.
Now, run a roscore in a separate terminal and get it out of the way. Set the rosparam use_sim_time to
true. Next, run the RViz configuration from this repo:
rviz -d /path/to/this/repo/config/asko_2024.rviz
Run the robot_description for both
vehicles in order to set up their TF trees (somehow they don't get set up from the rosbags):
roslaunch lolo_description lolo_description.launch
and
roslaunch usbl_tools boat_description.launch
You can go ahead and ctrl-c both nodes after they've been set up on RViz.
Finally, go ahead and rosbag play any of the datasets.
If all of the above went smoothly, you'll see the following RViz screen.
We have recently ported the datasets to ROS2 and have successfully worked with them on Humble. If you're interested in these feel free to raise an issue and we'll rework the instructions to guide ROS2 users as well.
We phrase this as a question since we don't particularly know how you can contribute towards this. We are always open to suggestions and collaborations. Our aim with making this datasets public is for them to be helpful for anyone who wants to do research in the field. Please feel free to raise any issues, even if it's just to leave a comment or start a conversation.
- The service boat's filtered navigation in the 2024-06-11-17-28-05 dataset is very noisy. It has strange biases and drifts from time to time. Challenging to work with. This is widely discussed in our publication.