Beyond DroneLibrary Multi-Domain Drone Modeling: Simulating and Flying a MultiBody Quadrotor in 3-D with Rumoca

Gallery project for the open challenge of the Rumoca workshop at the American Modelica Conference 2026 (CogniPilot, Georgia Tech, 2026-10-12). Companion entry: the rocket landing, ALSETLab/rumoca-rocket-landing.

Question. How far does Rumoca 0.10.2 get with a real Modelica.Mechanics.MultiBody quadrotor library, and can the library's own hover test fly in Rumoca's 3-D viewer instead of the licensed DLR Visualization library that the DroneLibrary normally uses?

Background: the DroneLibrary and its papers

• M. Podlaski, L. Vanfretti, H. Nademi and H. Chang, "UAV Dynamic System Modeling and Visualization using Modelica and FMI," The Vertical Flight Society's 76th Annual Forum & Technology Display, Virginia Beach, October 6-8, 2020. Publisher, author's copy.
• M. Podlaski, L. Vanfretti and D. Winkler, "DroneLibrary: Multi-Domain Drone Modeling in Modelica," Proceedings of the 15th International Modelica Conference, Aachen, October 9-11, 2023, doi:10.3384/ecp204551, author's copy.
• Models: ALSETLab/Modelica-Drone-3D-FMI.

The DroneLibrary models a quadrotor across domains: a MultiBody airframe, DC motors, rotors and blades, power electronics and batteries, sensors, and continuous, discrete and synchronous controllers. The 2020 paper covers its 3-D visualization and FMU export. The library's virtual-reality view uses the DLR Visualization library, which needs a commercial license.

What Rumoca adds

Topic	Papers (2020, 2023)	This project, with Rumoca 0.10.2
Compiler	Dymola	Rumoca simulates the unmodified Tests.simpleHoverTest (7613 equations) and Tests.motorTest
Cross-check	Dymola only	Rumoca agrees with OpenModelica to 0.22 mm in altitude over the first 4 s
3-D view	MultiBody animation in Dymola; VR with the DLR Visualization library (commercial license)	Rumoca's free, open-source Three.js viewer, live, from the MultiBody quaternion and rotor speeds
Support status	n/a	3 of 7 MultiBody models from two ALSETLab libraries simulate; blockers and work-arounds documented
Rumoca bugs	n/a	Reset and re-initialization failure reported with a reproducer, CogniPilot/rumoca#430

Starting model. DroneLibrary.Tests.simpleHoverTest: a chassis of four MultiBody arms, a cylinder and a point mass; four propellers, each with a DC motor, a revolute rotor and blades; a ramp altitude setpoint of 5 m and one PID driving all motors. The library is used unmodified: no equation, parameter or connection was changed.

Change. Everything is outside the library:

1. The Modelica Standard Library 4.0.0 is given to Rumoca as one root directory holding Modelica/, ModelicaServices/ and Complex/ side by side (0.10.2 does not resolve references across roots).
2. In that copy, the external "builtin" bodies of Modelica.Math.sin/cos/... are rewritten as algorithm y := .cos(u);, because the 0.10.2 runtime cannot execute external builtins, and two OpenModelica-specific calls in ModelicaServices are replaced.
3. A copy of the library gets an empty package.mo in the three folders that lack one, because the Rumoca CLI rejects such a tree as a source root.
4. A Rumoca scenario and a Three.js scene put the chassis pose from the MultiBody frame variables, and the four rotor speeds, into Rumoca's interactive viewer.

tools/build_rumoca_roots.py builds items 1 to 3 from OpenModelica's installed libraries and the cloned DroneLibrary; it never modifies the originals.

Results

Rumoca agrees with OpenModelica on the unmodified hover test

check/compare_om.py, figure check/compare_om.png, data check/compare_om.json. Rumoca 0.10.2 native simulator against OpenModelica 1.27 (dassl, tolerance 1e-6) on the same library files, first 4 s:

Signal	Max abs difference	Range over 4 s
Chassis altitude droneChassis1.bodyCylinder.frame_a.r_0[3]	2.2e-4 m	18.5 m
PID output (motor current command)	6.9e-4	136.8
Rotor speed propellerRev.rotor3_1.revolute.w	1.1e-4 rad/s	20.5 rad/s

Rumoca compiles the 7613-equation model in about 19 s and simulates 4 s in about 75 s on this laptop.

Both tools show the same behavior, which belongs to the library's test, not to either tool: the 50 kg chassis first sinks to 18.5 m below its start while the PID winds up, then overshoots to 70.7 m at 9.3 s (OpenModelica, 10 s run), and the loop then diverges. Over 40 s OpenModelica ends at 2202.0 m; Rumoca's live viewer showed 2202.3 m at 40.2 s.

The hover test in Rumoca's 3-D viewer: live and replay

Two ways to watch it, both in Rumoca's own viewer (PowerShell; the browser opens by itself once the viewer is up):

Mode	Start	What runs	Controls
Live (http://localhost:8090)	.\viewer\run_viewer.ps1	Rumoca simulates the unmodified simpleHoverTest (7613 equations) at 0.7 to 1.1 times real time; compiling takes one to two minutes	Camera button (C), T real-time toggle, F fullscreen, H heads-up display, Q quit
Replay (http://localhost:8092)	.\viewer\run_replay.ps1	Plays back a recorded 10 s Rumoca run of the same test, looping; starts in seconds	Replay from the start (R), Camera (C), T, Q

Both pages show a banner, LIVE or REPLAY, and a status line with time, altitude, set-point, rotor speed and PID output. The quadrotor is drawn from simple Three.js geometry; its position and attitude come from the MultiBody quaternion of the chassis cylinder, and the four rotor discs spin with the rotor speeds.

The live mode has no "fly again" or loop: Rumoca 0.10.2's reset fails on this model and ends the viewer (CogniPilot/rumoca#430). It also stops by itself at t = 81 s, after the test's PID loop has diverged.

The replay exists for that reason. check/record_replay.py records Rumoca's native simulation of the unmodified test (10 s, every 0.02 s, about 100 s of computing) into viewer/replay/hover_replay.json. viewer/DroneReplay.mo is only a replay clock that loops from 0 to 10 s. viewer/drone_replay_scene.js interpolates the recording at that clock; it is generated from drone_scene.js by tools/make_replay_scene.py. Nothing physical is simulated live in replay mode, and the page says so. Screenshots: results/viewer_live_controls.jpg and results/viewer_replay.jpg.

MultiBody support matrix

With the same two work-arounds, 3 of 7 MultiBody models from the two ALSETLab libraries simulate in Rumoca 0.10.2; 6 of 7 compile. Full table with exact diagnostics: multibody_support_matrix.md in the rocket repository.

Model	Equations	Compiles	Simulates	Blocker
DroneLibrary.Tests.motorTest	3466	yes	yes	none; matches OpenModelica to 3.7e-7 relative
DroneLibrary.Tests.simpleHoverTest	7613	yes	yes	none; this entry
DroneLibrary.Tests.propellerTest	4256	yes	no	algebraic projection does not converge at an event
DroneLibrary.Examples.DroneWithIdealPower.TestSystem	8005	yes	no	Solve-IR tensor fill extent

Rumoca 0.10.2 findings from this entry

• Modelica.Math external builtins block every MultiBody simulation ("external builtin function Modelica.Math.cos calls cos, which this runtime cannot execute"); the algorithm-body rewrite fixes it.
• Cross-root references do not resolve, so the MSL must be one root directory.
• The CLI rejects a source root containing a folder without package.mo (PKG-006), although the Python API loads the same model file from inside the tree.
• The Windows release binary does not ship the viewer's browser bundle: the page says "Missing Rumoca viewer shell". It looks for packages/rumoca-web/viz/viewer.html and packages/rumoca-web/vendor/three_viewer.js under the current folder. The external/rumoca-web-assets submodule (ALSETLab/rumoca-web-assets) holds both: the page from the v0.10.2 source and the bundle built with esbuild 0.28.1 from three 0.164.1, exactly as packages/rumoca-web/build.mjs does. The launch scripts run the CLI from that folder, which also downloads rumoca.exe on first use.
• The viewer's reset (R) restores the states but not the model clock, so time-based sources such as the library's Ramp setpoint do not replay after a reset.
• On this model the reset fails outright, and so does the start of a minimally changed variant; see the next section.
• A live viewer run stops at t = 81 s with "algebraic projection did not converge at event boundary" (worst scaled residual ratio 1.000001, target droneChassis1.bodyShape.frame_a.R.T[1,2]); by then the library's loop has diverged and the drone is above 2 km. Restart with viewer/run_viewer.ps1.

Reproducer: reset and re-initialization failure

Reported as CogniPilot/rumoca#430.

check/repro_reset/. Two symptoms with the same error:

1. In the viewer, pressing R on the running simpleHoverTest prints "reset: session reset failed ... initial homotopy continuation failed at lambda=0 ... algebraic projection did not converge at event boundary ... target=droneChassis1.bodyShape.frame_a.R.T[1,3]", and the viewer exits.
2. DroneHoverLoop.mo re-assembles simpleHoverTest with identical components and connections, plus one state flight_time (der = 1) that drives the 5 m ramp instead of time. It fails at session creation with the same error, in both the CLI and the Python API (check_loop_model.py, log check_loop_model.log), while simpleHoverTest itself initializes and runs.

How to run

Clone with the submodules: external/Modelica-Drone-3D-FMI (the original ALSETLab DroneLibrary) and external/rumoca-web-assets (ALSETLab/rumoca-web-assets, the viewer kit the Rumoca 0.10.2 Windows release binary needs). Prerequisites: Python 3.12 (py -3.12 or python) and OpenModelica 1.27 with the Modelica Standard Library 4.0.0 (the source of the patched MSL root).

git clone --recurse-submodules https://github.com/ALSETLab/rumoca-drone-library.git
cd rumoca-drone-library
.\setup.ps1
.\viewer\run_replay.ps1
.\viewer\run_viewer.ps1

setup.ps1 initializes the submodules, creates .venv, builds the source roots msl/ and libroot/ (tools/build_rumoca_roots.py) and downloads the Rumoca CLI into the viewer kit, checking its SHA-256. The live viewer opens http://localhost:8090 once compiling is done (one to two minutes). The comparison with OpenModelica, from the repository root:

.venv\Scripts\python.exe check\compare_om.py

It needs the OpenModelica result check/om/om_simpleHoverTest_res.csv (tracked), produced by omc om_simpleHoverTest.mos in check/om (OpenModelica 1.27).

Credits

DroneLibrary: M. Podlaski, H. Chang, L. Vanfretti (RPI) and D. Winkler (USN), BSD 3-Clause. Modelica Standard Library: Modelica Association, BSD 3-Clause. Viewer page and build recipe: Rumoca (CogniPilot, Apache-2.0). Tools: Rumoca 0.10.2, OpenModelica 1.27. Agent: Claude Code with Claude Fable 5.1 and, after 11:30 EDT, Claude Opus 5.5.

Lessons learned across both Rumoca projects: LESSONS_LEARNED.md.