ROS 2 Lyrical / Gazebo Jetty — DAVE migration verification

Can DAVE and its CUDA-free WGPU multibeam sonar (PR #44) run on ROS 2 Lyrical + Gazebo Jetty? DAVE is documented for Jazzy + Harmonic. Verified here on macOS (Apple Silicon, native) and Docker (Ubuntu 26.04).

This is a verification record, not a distribution. It establishes scoped build, launch, topic and service behavior. General numerical, acoustic, optical and hydrodynamic correctness is not established. On 2026-08-29 ten candidate patch groups addressed the defects reproduced and discriminated in that round's available Mac and Docker environments. The ninth closes the fifth-ROV sonar world-composition omission; the tenth sets an explicit valid ArduSub speedup after the official Gazebo plugin exposed an arm64 logger FPE. The upstream DAVE checkout and its installed workspace remain unchanged; those fixes live in patches/ and are backed by notes/results/remaining_defect_fixes_2026-08-29/, the open-gap revalidation and the external-stack validation. Later isolated Gazebo DVL, sonar startup and parameter_bridge ownership-cycle candidates are indexed in patches/README.md. Unlike the DAVE candidates, the four-file ros_gz_bridge ownership fix is now built in the user's actual ros_gz workspace and submitted as signed upstream PR #952. Hardware, external-account and broad scientific-validation boundaries remain open in notes/next-steps.md.

Results

Worlds	17/18 PASS-level, 1 PARTIAL in the world-level matrix — dvl_world.world remains PARTIAL because the distributed Homebrew/stock path is unchanged. An isolated exact-gz-sim10_10.4.0 candidate now removes the custom-DVL initialization race: the unmodified official world passes 20/20 without a hidden camera, but this is not installed or upstreamed
Direct world-model audit	A candidate patch gives all 18 distributed files unique internal <world name> values; Mac/Docker scoped regression checks pass. The earlier 14-name/3-duplicate-group result remains historical evidence, not the patched verdict — current evidence
Direct sonar check	The six-scene Docker matrix still separates incorrect distributed WGPU raw bins from the exact-N candidate. A 2026-08-31 fixed-seed/frame control then ran three fresh containers and captured nine raw plus nine point frames: each class was byte-identical within and across runs, and every raw peak was 4.013377666 m (error +0.013377666 m). Empty-cache software probing remained expensive (94.192–101 s versus about 5 s reused), so this is a deterministic correctness discriminator, not a production-performance or general-acoustic result — matrix · determinism/cache evidence
Direct DVL check	The isolated macOS exact-tag candidate still passes the official world 20/20, while the distributed Homebrew path remains PARTIAL. Separately, one Docker world instantiated all eight DAVE descriptors simultaneously and captured 20 valid messages from each (160 total): four beams, bottom lock, eight distinct nonempty frame IDs and descriptor-specific 8/12/7 Hz rates all passed. This closes bounded multi-device behavior, not physical calibration, mission endurance or upstream installation — Mac candidate · multi-device evidence
Direct ROV-model check	Patched Mac REXROV remains 7/7 in the configured state/sensor scope, and the isolated sonar/Heavy candidates retain their bounded output and control passes. The DAVE-sonar endpoint-dependent bridge defect (group shutdown exit -11 in 9/10) is closed in the reproduced parameter_bridge scope by the WeakPtr ownership candidate: bridge-first 20/20, group 10/10, direct ROS→GZ 20/20 and active camera/PointCloud 5/5 each. Cross-distribution ARM64 and emulated amd64 matrices pass. The candidate is now installed in the actual user ros_gz workspace, where it builds, passes 24/24 bounded directions, instantiates all 73 generated pairs in both directions, passes 73/73 generated payload assertions each way, and exits cleanly in the ordered 73-pair run; repeat payload 5/5 and lifecycle 11/11 also pass. Issue #951 and signed PR #952 are open. Separate macOS component/test-helper teardown remains: an ordered bridge_node run passes 3/3 topic + 1/1 service assertions but exits 139 on SIGINT. Native x86_64 hardware, Windows, hardware GPU and maintainer merge remain external. The exact tested DAVE image still has the documented IMU/magnetometer sensor-contract gaps — cycle-fix candidate · baseline bridge matrix · sensor contract
Current Docker image recipe	Fresh --no-cache PASS with rendered replay (2026-08-30). After official BuildKit cache pruning, the current Dockerfile completed from scratch in 66.917 minutes and produced a 6,260,137,751-byte arm64 image (53744d17f09d). Package and pinned-source checks pass. A real FreeRDP/xrdp login started Xorg/XFCE; the retained framebuffer shows Gazebo and QGroundControl with MAVROS connected: true, MANUAL and QGC Ready/Manual. The earlier cache image separately passed three headless control loops and a Windows App rendered replay — fresh evidence · cache-image controls
Direct Glider-model check	Docker RDP reran both Wiki launches and all 9 bridge topics. State/sensor 6/6 passes with the local IMU patch; cmd_thrust changes propeller velocity. A repeated integrated deadband control delivered 50/50 ROS true messages to Gazebo with 0 false, withdrawing the earlier one-shot asymmetry claim. Calibration, Mac stepping and long dynamics remain open — evidence
Direct Object-model check	Packaged, generic Fuel and copied-source descriptor paths pass. The candidate launch preflight now rejects a missing descriptor instead of printing false success on Mac and Docker. Fuel /1 remains a client/external reproducibility limitation, not an immutable pin — evidence
Direct USBL check	Candidate controls pass zero-noise, paused callbacks, moving targets and fractional 1539/1541 m delay on Mac/Docker. A 2026-08-31 Docker matrix additionally ran two transceivers and four transponders in namespaces A/B with reused IDs 1/2: A-only, B-only and concurrent phases delivered only the intended namespace, with maximum axis error 5.55e-17 m. Mission endurance and physical acoustic accuracy remain outside scope — candidate controls · namespace matrix
Direct ocean-current check	The 12 global services, two-depth/two-namespace model path and vehicle response pass in their controlled scopes. A 2026-08-31 Docker control added 400 Gauss–Markov samples with all three axes varying and a documented per-model tidal path: the no-tide model stayed fixed for 200 messages while the tide-enabled model produced 200 distinct X values over a 0.507469 m/s range. The earlier global-topic tide comparison is withdrawn as the wrong oracle because tidal oscillation is applied downstream per model — depth/vehicle evidence · noise/tide evidence
Direct underwater-camera check	The semantic R/B candidate remains validated. A separate 2026-08-31 Docker isolation matrix exercised all six attenuation/background tags plus controls simultaneously, captured 10 frames per camera and matched every analytic centre-pixel prediction within 0 LSB; each semantic R/G/B tag affected the intended output channel. General underwater optical accuracy across materials, scattering and real water remains unverified — candidate fix · six-tag matrix
Direct SeaPressure check	Candidate controls pass Pa/Pa² units, sign, saturation, noise, rate, variance, frame and topic behavior on Mac/Docker. A 2026-08-31 Docker run kept seven devices active to about 200 simulated seconds; 2,000 noisy frames had mean 101325.214 Pa, standard deviation 119.914 Pa for target 123 Pa, exact variance 15129 and all unique values, while 2/10 Hz controls passed. This is bounded synthetic contract/statistical validation, not real-sensor calibration or mission endurance — candidate controls · long/multi evidence
Direct Spherical Coordinates check	The candidate patch keeps the independent valid-input oracle result, rejects NaN/out-of-range requests, returns explicit failure status, responds while paused, fails safely without world spherical configuration and fixes plugin discovery on Mac/Docker. Adding success fields is an interface change — evidence
Direct Model Plugin check	Interpolation and vehicle motion had already passed separately. A simultaneous fixed-depth Docker control now verifies depth-dependent application and two namespaces: 5 m → 0 m/s and Δx 0; 15 m → about 0.75 m/s and Δx 2.400436 m over 6 s. Coefficient and real-ocean accuracy remain outside this result — evidence
Sonar cost	RTF 0.5243 against a 0.9974 no-sonar control — 1.90x, not the 4.5x reported earlier
Largest win	update_rate 30 Hz → 2 Hz removes 75% of that cost, with no code change
Build trap (pre-fix history)	The old guides set no CMAKE_BUILD_TYPE, so nothing received an -O flag. The candidate patch adds Release build commands to both guides; the measured sonar-world change was 0.218 → 0.438 (n=1)

Measurement detail and how these numbers replaced the earlier ones: notes/sonar-performance.md.

Build

extras/build-dave-lyrical-linux.sh     # Ubuntu 26.04 — complete build path; prints the validated launch command
extras/build-dave-lyrical-macos.sh     # macOS Apple Silicon — see the platform notes for one gap

Platform-specific caveats: extras/README.md.

Both refuse to finish if the sonar was compiled without an -O flag. That is not caution: this workspace built unoptimised for a month, and the sonar figures taken in that time were about half what they should have been. The measured effect is one condition of one world — RTF 0.218 → 0.438 on dave_multibeam_sonar at its shipped sensor configuration, n=1 — so treat 2.01x as that result rather than as a workspace-wide multiplier. Other packages were not re-measured.

Building by hand instead, the one thing that must not be omitted:

colcon build --cmake-args -DCMAKE_BUILD_TYPE=Release

# expect -O3, not an empty result
grep -o '\-O[0-3s]*' build/multibeam_sonar/compile_commands.json | sort | uniq -c

Without it colcon leaves the build type empty and nothing gets an -O flag. On the sonar world that is RTF 0.2180 against 0.4380. Every performance figure here dated before 2026-08-05 was taken without it.

The full procedure as prose, with the reasoning for each argument: notes/setup/reproduction.md. For Docker, docker/.

Documentation

Everything is under notes/ — start from its index.

notes/what-we-got-wrong.md	claims that turned out false, and how each was caught. Read this before trusting any number here
patches/	what was changed — eighteen retained patches; the ten 2026-08-29 candidates and two 2026-08-30 isolated candidates include validation and dependency order
notes/verified-demos.md	what each PASS rests on
notes/known-issues.md	49 entries, including open, resolved and withdrawn findings, with cause and workaround
notes/progress-log.md	112 dated rows: what was done each day, and what later turned out wrong
notes/setup/reproduction.md	annotated build and launch procedure
notes/validation_matrix.csv	every world and vehicle, with its label
notes/upstream/submit/	eleven issue records prepared (ten for IOES-Lab/dave, one for gazebosim/ros_gz); the bridge-lifecycle report is issue #951 with signed PR #952, the other ten remain unfiled, and one is explicitly historical/do-not-file
docker/	Docker image build, RDP desktop, verification commands
notes/next-steps.md	what is still open
notes/wiki/	corrections applied to the DAVE documentation, and what was deliberately left out

Test labels

Label	Meaning
SMOKE PASS	Launches and stays alive for the test window. Topic data was not checked
FUNCTIONAL PASS	Expected topics, services or sensor data were actually read
PARTIAL	Runs and produces some output, but has a confirmed functional or performance problem
NOT AUTOMATED	Not reachable through the current headless path. Not a claim that it cannot be tested

Environment

	macOS (native)	Docker
OS	macOS 15.7.3, Apple Silicon (M2)	Ubuntu 26.04 Resolute (arm64)
ROS 2	Lyrical (source build)	Lyrical (apt, ros-lyrical-desktop)
Gazebo	Jetty (Homebrew)	Jetty 10.5.0 in the fresh 2026-08-30 image (apt vendor build); older retained runs used 10.4.0
Python	3.14	3.14
Rendering	Metal (real hardware)	Vulkan llvmpipe (CPU software renderer) — no /dev/dri passthrough. The distributed software-WGPU ordering still reproduces the OGRE2 sample-texture/null-memcpy crash; an isolated deferred-backend candidate preserves ogre2 output across 20/20 cold starts and 3/3 Heavy-control repetitions
Sonar compute	WGPU on the Metal adapter; CPU backend also run explicitly	CPU fallback and forced CPU are validated. An isolated candidate also runs WGPU on Docker llvmpipe and Mac Metal, but it is not installed or upstream. The exact-N correctness discriminator is too slow on fresh llvmpipe (88.2–115.3 s first probe). Docker hardware WGPU/CUDA remains untested

Pinned source revisions and the migration patch: notes/patch-and-pinned-commits.md.

References

• Choi, W.-S., "Ray-Based Physical Modeling and Simulation of Multibeam Sonar for Underwater Robotics in ROS-Gazebo Framework," Sensors 2025, 25(5), 1516. 10.3390/s25051516 — the method PR #44 implements
• Choi, W. et al., "Physics-Based Modelling and Simulation of Multibeam Echosounder Perception for Autonomous Underwater Manipulation," Frontiers in Robotics and AI, 2021. 10.3389/frobt.2021.706646 — the earlier raster-based work this sonar lineage descends from
• DAVE documentation · ROS 2 Lyrical Luth