uw-vio-dvl

Underwater stereo-inertial-pressure VIO (GTSAM fixed-lag) with a tightly coupled DVL velocity factor whose gating and re-lock logic keep DVL dropouts from corrupting the estimate. Research code; never run on a vehicle.

20 s camera blackout in Stonefish: vision-only vs with DVL	Top-down trajectory, NTNU traj_03
physically simulated DVL, single run, aligned on the pre-blackout segment: vision-only error grows to 8.8 m, with the DVL it stays under 0.5 m (sim details)	mocap ground truth vs 20 s blackout without DVL and with simulated DVL

Results

NTNU underwater 2023, 6 trajectories with mocap (mean ATE SE3 with hand-eye, RPE over 5 m). The DVL is simulated from mocap (noise, scale bias, ~2.5 deg mount error, dropouts, outliers):

configuration	mean ATE (m)	mean RPE (%)	worst traj (m)
VIO + pressure, cameras normal (base)	0.401	6.75	0.642
+ DVL (sim), periodic / long / burst dropouts, outliers	0.393-0.400	6.71	0.658
20 s camera blackout, no DVL	3.876	33.81	10.711
20 s camera blackout, DVL (sim)	0.663	7.31	0.936
20 s camera blackout, DVL also lost for 7 s inside it	0.888	9.39	1.527

Baselines on the same bags: Basalt 1.10 m (in-air stereo extrinsic), SVIn 1.80 m, ORB-SLAM3 initialises on 3/6, OpenVINS 3/6.

Stonefish seabed (physically simulated DVL with natural bottom-lock loss, exact ground truth, single runs):

clear water	ATE (m)
no blackout	0.30
20 s camera blackout, no DVL	3.69
20 s camera blackout, with DVL	0.32

In turbid water vision-only diverges and the DVL keeps it at 4.23 m.

Quick start

source /opt/ros/jazzy/setup.bash    # from bash, not zsh
colcon build --base-paths uw_ct_vio --cmake-args -DCMAKE_BUILD_TYPE=Release \
  -DGTSAM_DIR=$GTSAM_PREFIX/lib/cmake/GTSAM -DGTSAM_UNSTABLE_DIR=$GTSAM_PREFIX/lib/cmake/GTSAM_UNSTABLE -DBUILD_TESTING=ON
colcon test --packages-select uw_ct_vio && colcon test-result --verbose
run_bag --bag <ntnu>/traj_01/traj_01_ros2 ... --out traj_01.tum   # full flag list in docs/USAGE.md

More

• docs/RESULTS.md: estimator overview, the full results table, all figures (error vs time, per-trajectory ATE, dropout robustness, simulated DVL example) and the Stonefish simulation.
• docs/USAGE.md: build dependencies, the NTNU run example, the DVL simulator, evaluation, the benchmark runner and what is not validated.
• sim/README.md: Stonefish benchmark details and caveats.
• DESIGN_DVL.md (DVL design, as-built notes), DESIGN_V2.md and DESIGN_PHASE2.md (estimator history and ablations).

Not validated: real DVL data, the --dvl-topic reader, DVL extrinsic / scale / time-offset estimation, vehicle runs. Licence: none yet (choose before publishing).