Status: CURRENT RECIPE FRESH-BUILT AND RENDERED-RUNTIME CHECKED — 2026-08-30.
After official BuildKit cache pruning, the full current Dockerfile completed with `--no-cache`
in 4015.02 seconds (66.917 minutes), return code 0. The arm64 image is
`sha256:53744d17f09dd2489d6be3eedbfae19377b9013211f1cfa3c109c70cfc67d955`
(6,260,137,751 bytes), with Ubuntu 26.04.1, ROS 2 Lyrical and Gazebo Jetty 10.5.0.
DAVE, multibeam, MAVROS/mavros_msgs, ArduSub, QGC, the official ArduPilot Gazebo plugin
and all four pinned source revisions passed image checks.
A real FreeRDP/xrdp login to this fresh image started Xorg :10 and XFCE. The retained
1280×900 framebuffer shows Gazebo and QGroundControl rendered together; MAVROS reported
connected:true and MANUAL, while QGC visibly reported Ready/Manual. QGC still requires
QGC_NO_SYSTEM_GLIB=1. This fresh-image run used FreeRDP SDL. A separate retained replay on
the earlier exact cache image used Windows App, so those client claims are not conflated.
The earlier exact cache image remains the evidence for three bounded baseline/Heavy/
Heavy-multibeam headless control loops. The combined fifth-sonar candidate remains
backend-dependent: software WGPU/llvmpipe exits 139, while forced CPU publishes sonar and
completes MANUAL arm/control/disarm.
Fresh current-recipe evidence (2026-08-30): the Dockerfile SHA-256 was
595ad3d51c5d0f639eac465be617cf103bf3407e29ed5d7697b7435723ac9532 at Git
e05070a. The build began 11:13:38 KST and ended 12:20:34 KST after 66.917 minutes.
docker image inspect reports 6,260,137,751 bytes, arm64/Linux and final user root.
The verifier confirms ROS 2 Lyrical, Gazebo 10.5.0, DAVE/multibeam/MAVROS package
prefixes, required binaries/plugins/QGC files and pinned DAVE/geographic_info/ArduPilot/
ardupilot_gazebo commits. The only build warning recommends JSON-array CMD form and did
not fail the image. A real FreeRDP/xrdp/Xorg/XFCE replay then captured the rendered desktop
and a Gazebo+QGC vehicle-connected framebuffer. Evidence is under
final_gap_validation_2026-08-30/docker_no_cache_build/.
Earlier exact-cache evidence: a cache-assisted build completed in 44.86 minutes and
produced image af9586fa8045; that image passed package/pin checks and bounded baseline,
Heavy and Heavy-multibeam headless control loops. Its separate Windows App RDP replay is
retained under docker_exact_rdp/.
Provenance (2026-07-18): lyrical.arm64v8.dockerfile in this folder was clean-built
(--no-cache) end to end on 2026-07-18, including the commit SHA pinning, .bashrc
source-order fix, the CA-bootstrap fix described under Known limitations, and a from-source
mavros build (ros-lyrical-mavros isn't published via apt yet — see the main
README.md Known issues) — image tag lyrical-sim:jetty-rdp-pr1-ca-fix.
Build took 51m 22s, produced a 21.9GB image, and an idle running container (no active
RDP session or Gazebo demo) used 62.66MiB RAM (docker stats) — a floor, not a
representative figure under load. Under an active demo workload (2026-07-20, same image,
representative smoke test below running headless via docker exec -d), docker stats held
steady at ~1008MiB (8.44% of the 11.67GiB container memory limit) and ~1.2–1.5% CPU
across 5 samples over ~20s once the sonar plugin finished loading — roughly 16× the idle floor,
and consistent with a ps aux breakdown inside the container (gz-sim-main ~836MB RSS / ~15.7%
of one core, parameter_bridge ~90MB). Confirmed in that build/run: all 36 build steps completed,
ROS_DISTRO=lyrical and ros2 resolved inside the container, ros2 pkg prefix resolved
dave_demos and multibeam_sonar_system, and ros2 pkg list listed mavros, mavros_extras,
mavros_msgs, mavros_examples. This particular build/run round (2026-07-18) is clean build +
ROS environment/package presence validation — it did not itself launch the representative demo.
Corrected 2026-07-23 (was previously stated inconsistently): the representative-demo /
RAM-CPU-under-load measurement above was run against this exact image tag
(lyrical-sim:jetty-rdp-pr1-ca-fix), two days later on 2026-07-20 — see the main README's
Progress Log entry for that date, which names the image tag explicitly. The
world/vehicle/performance/stability rows in the table below, separately, were validated on
2026-07-22/23 against the same Dockerfile lineage running as the lyrical-theme-test
container (a different, later container from the same Dockerfile, not this exact image tag) —
see the table for current status and the main README for full detail. Real
RDP login to an XFCE desktop was validated against the same Dockerfile lineage on 2026-07-17 and
again during the 2026-07-18 --privileged test below; it was not re-clicked-through on this
exact rebuild since nothing in the RDP/XFCE stack changed. This replaces an earlier draft (with
separate entrypoint.sh/startwm.sh files) that was never actually built.
The lab's advisor has indicated this project will eventually move off the DAVE name to
something new — not yet decided. Until that's settled:
- External/display names (Docker image tag, container name, hostname, shell prompt) use a
neutral placeholder,
lyrical-sim, instead ofDAVE. This is what was actually used in the validated build (lyrical-sim:jetty-rdp, containerlyrical-sim, hostnamelyrical-docker, promptdocker@lyrical_docker:~$). DAVEis still used to refer to the underlying codebase and its existing ROS packages (dave_demos,dave_interfaces,dave_worlds, thedave_wsworkspace path, etc.) — those are not renamed here; that's a separate migration once the new project name is final.- Treat
lyrical-simas a placeholder, not a committed project name.
docker/
├── lyrical.arm64v8.dockerfile image definition — single self-contained file
└── README.md this file
No separate entrypoint.sh/startwm.sh: startwm.sh is generated inline (RUN printf ...)
directly into /etc/xrdp/startwm.sh inside the image, bypassing the default Debian Xsession
lookup chain, and the multi-service startup (dbus, sshd, xrdp-sesman, xrdp) is a single
shell-form CMD at the end of the Dockerfile — no separate entrypoint script needed.
Run from the repository root:
docker build --no-cache -f docker/lyrical.arm64v8.dockerfile -t lyrical-sim:jetty-rdp .--platform linux/arm64 is implied by the arm64v8/ubuntu:26.04 base image; this Dockerfile
targets Apple Silicon / arm64 only — it has not been adapted or tested for amd64.
The image's final USER is root (needed for the RDP/xrdp CMD), so bash -lc does not
auto-source /home/docker/.bashrc — the ROS/DAVE/mavros environment lines live there, not in
root's shell rc. Source the underlays explicitly instead of relying on login-shell sourcing.
ros2 --version is also not a real ros2 CLI flag; check $ROS_DISTRO and which ros2 instead.
docker run --rm lyrical-sim:jetty-rdp uname -m # expect: aarch64
docker run --rm lyrical-sim:jetty-rdp lsb_release -a # expect: 26.04 Resolute
docker run --rm lyrical-sim:jetty-rdp bash -lc \
'source /opt/ros/lyrical/setup.bash && echo "$ROS_DISTRO" && which ros2'
docker run --rm lyrical-sim:jetty-rdp bash -lc \
'source /opt/ros/lyrical/setup.bash && gz sim --versions'
docker run --rm lyrical-sim:jetty-rdp bash -lc \
'source /opt/ros/lyrical/setup.bash && \
source /home/docker/dave_ws/install/setup.bash && \
ros2 pkg list | grep "^dave_"'
docker run --rm lyrical-sim:jetty-rdp bash -lc \
'source /opt/ros/lyrical/setup.bash && \
source /home/docker/mavros_ws/install/setup.bash && \
ros2 pkg list | grep "^mavros"'docker run -d \
--name lyrical-sim \
--hostname lyrical-docker \
-p 127.0.0.1:3393:3389 \
lyrical-sim:jetty-rdp(Port left-hand side is arbitrary — pick one that doesn't collide with any container already
using 3389 locally.) 127.0.0.1: binds the published port to localhost only, matching the
warning below — omitting it would expose RDP on every network interface. Connect with any RDP
client (Microsoft Remote Desktop, etc.) to localhost:3393, user docker, password docker —
local development default only; change the password before running this anywhere
network-reachable. A successful login reaches an XFCE desktop with shell prompt
docker@lyrical_docker:~$.
--privileged is not required for container startup and XFCE/xrdp login — tested 2026-07-18
by running the image with no --privileged flag and no extra --cap-add (docker run -d --name priv-test --hostname lyrical-docker -p 127.0.0.1:3396:3389 lyrical-sim:jetty-rdp-pr1-ca-fix):
RDP login reached a usable XFCE desktop with no capability errors. xorgxrdp doesn't need host
GPU device access here (this image uses llvmpipe software rendering, no /dev/dri passthrough).
Dropped from the example above accordingly. Scope caveat: this only re-validates container
startup and RDP/XFCE login without --privileged — it has not been re-checked across every
Gazebo world, device-access path, or vehicle/sensor combination, so treat "not required" as
scoped to what was actually re-tested, not as a blanket clearance for every workload in this image.
A representative smoke test (headless, no RDP needed):
docker exec -it lyrical-sim bash -lc \
"source /opt/ros/lyrical/setup.bash && source \$DAVE_UNDERLAY/install/setup.bash && \
ros2 launch dave_demos dave_sensor.launch.py namespace:=blueview_p900 \
world_name:=dave_multibeam_sonar paused:=false x:=5.8 z:=2 yaw:=3.14 \
compute_backend:=wgpu gui:=true headless:=true"| Status | |
|---|---|
| Docker build provenance | 2026-07-18 clean baseline PASS; 2026-08-30 current recipe fresh --no-cache PASS. Current run: 66.917 min, return code 0, image 53744d17f09d, package/pin verification PASS |
| Current Dockerfile image | Fresh build and rendered runtime PASS in tested scope. Fresh image 53744d17f09d started a real FreeRDP/xrdp/Xorg/XFCE session and rendered Gazebo+QGC with MAVROS connected/Manual; the earlier cache image af9586fa8045 separately retains three bounded headless vehicle controls and a Windows App rendered replay |
| Container startup | PASS |
| xrdp connection, XFCE session (real RDP login, prompt confirmed) | PASS |
| ROS 2 Lyrical / Gazebo Jetty environment | PASS |
Representative REXROV launch + ros_gz bridges |
PASS |
WGPU/Rust sonar packages (wgpu_vendor, multibeam_sonar, multibeam_sonar_system) build |
PASS |
| ArduSub SITL build | PASS |
| mavros build and ArduSub bridge | FUNCTIONAL PASS in exact current image. Source-built packages resolve; all three BlueROV variants received Gazebo JSON, connected MAVROS 4/4 and completed one headless arm/control/disarm run |
| RAM/CPU under an active demo workload | MEASURED — short observation window: ~1008MiB (8.44%) / ~1.2–1.5% CPU, steady over 5 samples across ~20s (2026-07-20); idle floor is 62.66MiB. Not a long-duration measurement — see the separate long-duration stability test row |
| USBL | FUNCTIONAL PASS WITH REQUIRED WORKAROUNDS (direct Mac+Docker revalidation 2026-08-27). Common mode returned both tutorial transponder IDs and individual channels returned only the selected ID in both spherical and Cartesian streams. Largest retained positive-sigma static-coordinate axis error: 0.000258 m. Three independent limitations remain: (1) literal sigma=0 returns finite data on macOS/libc++ but aborts Docker/libstdc++ on the first ping (exit 134), so the 0.0001 world patch remains required for portability; (2) paused simulations expose endpoints but produce no payload because spin_some() is gated on !paused; (3) the old Wiki generic sensor launcher also tries to spawn nonexistent description/usbl/model.sdf. The verified command is ros2 launch dave_demos dave_world.launch.py world_name:=usbl_tutorial. Static routing/geometry only, not general acoustic/travel-time accuracy — see Known issues, Verified demos, and direct evidence |
| 18-world validation matrix | 18/18 PASS-level, 0 PARTIAL (2026-08-07) — 13 SMOKE PASS (process-level liveness), 5 FUNCTIONAL PASS (real topic/service/sensor data read back). The last two PARTIAL rows were reclassified when their stated grounds stopped reproducing: dave_multibeam_sonar → FUNCTIONAL PASS once it ran under the authorised-X-display configuration and published PointCloud2, and dave_ocean_waves_sonar_integrated → SMOKE PASS once its RTF and CPU-climb grounds were both refuted. The dave_world.launch.py headless gap that blocked all 3 manipulation worlds is fixed (patches/dave_world_launch_headless_fix.diff) — still Mac-only, not applied inside Docker. See notes/verified-demos.md and notes/validation_matrix.csv. Earlier 16/18 wording in Superseded figures |
Worlds previously documented as sonar-demo-branch-only |
The separate-branch requirement was investigated and found incorrect for this checkout (2026-07-22) — both worlds only need multibeam_sonar_system, already present. Both are now PASS-level: dave_ocean_waves_sonar SMOKE PASS, dave_ocean_waves_sonar_integrated SMOKE PASS (reclassified 2026-08-07). Neither has had its sonar output topics echoed — only /world/.../stats — which is why both are SMOKE rather than FUNCTIONAL. See notes/verified-demos.md |
| Vehicle/sensor/control coverage (not every Cartesian combination) | REXROV and Slocum Glider state/sensor scopes are FUNCTIONAL; BlueROV2 and Heavy sensor scopes remain partial because their configs bridge an undeclared magnetometer. Baseline, Heavy and Heavy-multibeam control loops are FUNCTIONAL in one exact-image run each, and the isolated fifth-sonar candidate is FUNCTIONAL on Mac. The later combined Docker derived-image run is FAIL/PARTIAL: after 60053 ms WGPU startup and sonar configuration, Gazebo began a stack trace and JSON/MAVROS, PointCloud2 and control were not reached. DVL/camera/ocean-current are FUNCTIONAL in their stated scopes; SeaPressure and USBL current candidate verdicts are in Verified demos. |
| Quantitative performance benchmark | PASS for dave_ocean_waves/usbl_tutorial, clean rerun 2026-07-27: dave_ocean_waves RTF 0.170 (Docker) / 0.376 (Mac); usbl_tutorial RTF 0.687 (Docker) / 0.646 (Mac). Reported as separate environments per world, not a single "N× faster" claim. No valid Docker RTF exists for dave_multibeam_sonar — the figures previously quoted here were withdrawn (see Superseded figures), and the world has not been re-measured under the X-display configuration that makes it run. Mac figures for that world are in notes/sonar-performance.md. See notes/next-steps.md and notes/bench_results/ |
| Long-duration stability | PASS (2026-07-29) — a genuinely clean 4-hour run completed: 120 samples at an unbroken 2-minute cadence from elapsed=2min to elapsed=240min with no gaps, Outcome: SURVIVED full planned duration, RSS 932 → 1253 MiB (+12%, memory-growth heuristic OK). This is the first 4h run that is simultaneously current-script, uncontaminated and gap-free — it closes the item that had been open since 2026-07-27. Raw data committed at notes/stability/4h_clean_2026-07-29/ and re-verified on the committed copy. Earlier attempts in Superseded figures |
arm64v8/ubuntu:26.04(this tested minimal image) ships withoutca-certificates, so any HTTPS apt source failed before the first package installed, and the defaulthttp://mirror also failed (connection timeout). Confirmed Docker's own network/HTTPS path was fine (docker run --rm curlimages/curl:latest -I https://ports.ubuntu.com/ubuntu-ports/dists/resolute/InRelease→200 OK), which isolated the cause to the missing CA bundle in the base image rather than a network/proxy problem. Fixed by copying the CA bundle from a digest-pinnedcurlimages/curlimage into/etc/ssl/certs/before any HTTPS apt operation (the exact path was confirmed against that digest, not assumed — the Dockerfile also asserts the copied file is non-empty and fails the build immediately if that ever stops being true), forcing anyhttp://ports.ubuntu.comapt source tohttps://, and addingca-certificatesplus an explicitAcquire::https::CaInfo=/etc/ssl/certs/ca-certificates.crtto both the firstapt-get updateand the firstapt-get install— setting it onupdatealone let the index download succeed but the followinginstallstill failed with an SSL error. A clean (--no-cache) build with this fix in place completed successfully on 2026-07-17, followed by a real RDP login. This is scoped to this tested minimal image, not a claim that every Ubuntu 26.04 base image lacks CA certificates.- Ubuntu 26.04 GNOME 50 is Wayland-only in the tested image;
xorgxrdpis X11-only, so it can't start a GNOME session here. XFCE still ships a full X11 session, so that's what this image's RDP desktop uses. This is an observation about this tested image, not a universal claim that GNOME remote desktop is impossible on every Ubuntu 26.04 configuration. - Docker's
llvmpipeis CPU software rendering, not hardware GPU acceleration. The Mac Metal comparison elsewhere in this repo is a real Apple M2 hardware GPU backend — the two numbers are environment-specific observations, not a controlled benchmark (OS, native vs. container, and GPU vs. CPU renderer all differ at once). CMDis currently shell-form, which produces aJSONArgsRecommendedwarning during build (not a failure). Switching to anentrypoint.sh+ JSON/exec-formCMDis possible but wasn't done since it risks changing how the multi-service startup (dbus, sshd, xrdp-sesman, xrdp) currently works — tracked as a follow-up, not fixed blindly.ShaderParamSDF warning (ShaderParam plugin element not defined in SDF/Copying plugin as children of sdf) seen during REXROV spawn; robot spawn andros_gzbridge creation completed successfully afterward, so this is a non-fatal known warning, not an error.- A "Stack trace (most recent call last) in thread N:" line appears in
gazebo-1's log right after the multibeam sonar plugin finishes loading (seen 2026-07-20, during the RAM/CPU-under-load measurement above). This did not crash the process:ps auxinside the container showedgz-sim-mainstill alive and actively consuming CPU (~15.7% of one core) more than 7 minutes after the trace appeared, and thedocker statsRAM reading was stable across all 5 samples. Also present in the same log:error: XDG_RUNTIME_DIR is invalid or not set in the environment— non-fatal, the WGPU sonar backend still selectedllvmpipeand compiled its pipelines successfully afterward. Root cause of the stack trace itself not yet investigated; flagged here as a known non-fatal log artifact, not confirmed benign for every world/workload. - Rebuilding any workspace (
dave_ws,mavros_ws, the ArduSub SITL build) as the runtimedockeruser fails (found 2026-07-20, while launching demos over RDP). All workspaces are compiled once, asroot, during the image build itself — thedockeruser never needs to runcolcon buildat all; sourcing the already-builtinstall/setup.bashand launching directly is sufficient (and is what every command in this file does). If a rebuild is attempted anyway asdocker, three separate issues surface in sequence: (1)cargo(needed bywgpu_vendor) only exists under/root/.rustup/..., which thedockeruser can't read — fix by installing rustup separately fordocker(curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh -s -- -y); (2) the existing workspace directories under/home/dockerareroot-owned from the image build, socolcon buildfails withPermission denied— fix withsudo chown -R docker:docker /home/docker; (3)ardupilot_sitl'swafstill fails withModuleNotFoundError: No module named 'imp'even after both of the above, the same Python 3.14/impincompatibility already tracked in the main README.md Known issues — the shim applied during the image build isn't present/effective in this interactivedocker-user shell. Net effect: don't rebuild asdockerunless you're intentionally testing uncommitted source changes; for running the existing demos, skipcolcon buildentirely.
Nothing in this section describes the current state. It is kept so that the record shows when a number changed and why. Do not quote from it.
Withdrawn: the Docker RTF ~0.0018 / ~0.0008-0.0077 figures for dave_multibeam_sonar.
They were sampled from /world/oceans_waves/stats, which is not that world —
dave_multibeam_sonar.world declares <world name="default"> — while a dave_ocean_waves
stability run was executing in the same container. Withdrawn 2026-08-03, together with the
"Docker is ~123x worse than Mac" comparison built on them.
Superseded: RTF ~0.03 for dave_ocean_waves_sonar_integrated (a 20 s spot sample; Mac
measurements under the corrected method give 0.2241 / 0.2197). The Mac 4.5x sonar cost
(now 1.90x — the earlier figure came from an unoptimised build with a settle criterion that was
timing the startup window). The 32%→47%→69% CPU climb (not reproduced in two isolated runs).
Superseded wording: this page's table previously read 16/18 PASS-level, 2/18 PARTIAL, and
its stability row read PARTIAL ... a proper rerun is still needed. Both were true when written
and are not now — see the rows above.
Full accounting: notes/what-we-got-wrong.md.