SM-Bello/C172P-Grounded-JSBSim-Airborne-Trim-Failure-Negative-Result

c172p Grounded

A Negative Result: JSBSim's c172p Could Not Be Trimmed for Level Flight

---

Why this dataset exists

Most published aerospace ML/control work only shows what worked. This one doesn't.

This is a negative result from the early stage of the PHI-CTRL project (Physics-Hybrid Integrity Control — a fault-tolerant flight control architecture). Before the project settled on the F-16A as its plant model, the original plan was to build and validate the whole fault-injection / hybrid-compensator pipeline on JSBSim's general-aviation c172p (Cessna 172P) — a smaller, slower, cheaper-to-simulate airframe that would have been a more approachable starting point.

It didn't work. Six independent control strategies, six failures to hold level flight. This dataset is the full record of that: the run scripts, the telemetry, the plots, the console logs, and the environment provenance — published as-is, not cleaned up or explained away, because a documented failure is still useful data. It's also a warning label for anyone else about to build serious control work on top of c172p: JSBSim's own upstream model file marks it explicitly as BETA — "This aircraft model probably will not fly as expected. Use this model for development purposes ONLY!!!" — and this dataset is six independent, structurally different pieces of evidence for why that warning should be taken literally.

---

The six attempts

#	Strategy	Result (this run)
01	Native trim search (do_trim(1))	do_trim() raised an explicit "Trim Failed" exception. No usable trim point returned; throttle/elevator both fell back to 0.0.
02	Forced initial conditions, fixed controls, no trim solver	Sustained descent, -211.6 ft over 30 s
03	Engine spool-up window before control engagement	RPM stabilized cleanly at t=2.01 s, but sustained descent still followed once control engaged: -355.1 ft over 30 s
04	Gain retuning (cascaded PID, retuned for c172p's dynamic scale)	Sustained descent, -121.3 ft over 30 s (least bad of the six, still not level)
05	TECS (Total Energy Control System)	Sustained descent, -138.0 ft over 30 s
06	Hard rate limiting (elevator + throttle) on top of Attempt 04's PID	Sustained descent, -127.9 ft over 30 s

None of the six held altitude. All six show continuous, unarrested descent rather than a stable equilibrium, regardless of control strategy sophistication — attempts 04–06, the most engineered of the six, are marginally better than 02–03 but still fail the basic test of holding altitude.

What this suggests

The consistency of failure across six structurally different control strategies — including one (01) that bypasses any custom logic entirely and relies purely on JSBSim's own solver — points to the problem sitting in the aircraft definition itself (c172p.xml, upstream BETA status), not in any particular control law tested here. Attempt 01's explicit solver failure is the strongest single piece of evidence for this: if JSBSim's own iterative trim solver cannot find an equilibrium for this airframe at 1000 ft / 90 kts, that is independent of anything downstream of it.

No dataset of clean, controllable flight episodes could be produced from this airframe. That's the negative result in full: not "the control law needs more tuning," but "there was no stable operating point to build a fault-tolerant control dataset around in the first place."

---

Where the project went instead

Rather than keep fighting an airframe definition that JSBSim's own maintainers flag as unreliable, PHI-CTRL pivoted to the F-16A model, which trims cleanly and gave the project a controllable, fault-injectable platform that actually matched the original intent — a plant stable enough that injected actuator faults, not simulation artifacts, are what the control architecture is being tested against.

That pivot produced the rest of the PHI-CTRL project:

Repository	What's there
Source code — Sm-bello/PHI-CTRL	The full hybrid control architecture (baseline, MRAC, MMAE bank, observer, PPO residual), verification pyramid, and this failure documented in docs/
F-16A fault telemetry — PHI-CTRL-F16-Fault-Recovery-Telemetry	The dataset this one couldn't produce — clean trimmed flight, then 160 episodes of elevator-effectiveness faults across γ ∈ {1.0, 0.8, 0.6, 0.5}
Trained models — PHI-CTRL-F16-Models	Residual PPO policy and CNN-BiLSTM effectiveness twin, trained on the F-16A data above
Archival record	DOI: 10.5281/zenodo.22218809 — frozen snapshot of the v1.0.0 release

This dataset is the reason that pivot happened. It's kept public and separate, rather than quietly deleted, because the six-way failure is itself a useful data point for anyone else evaluating c172p for serious work.

---

Open question — can this actually be fixed?

This is left as an open question, not a closed case. Six attempts is not exhaustive, and it's entirely possible the problem is fixable with the right domain knowledge that wasn't applied here. Things that weren't tried and might change the outcome:

• Rebuilding or hand-correcting the c172p.xml aerodynamic/propulsion tables rather than treating them as fixed
• Cross-checking against FlightGear's own c172p flight model or a different open-source C172 FDM for a sanity baseline
• A from-scratch trim search with a wider initial-condition grid and looser convergence tolerance than JSBSim's default do_trim()
• Reduced-order linearization around candidate trim points to check whether an equilibrium exists near where the solver gives up, even if the full nonlinear solver can't converge to it
• Filing this as a reproducible bug report against the upstream jsbsim-aircraft model itself, since a solver-confirmed "Trim Failed" on a shipped default airframe is arguably a defect worth the maintainers' attention

If you work on flight dynamics, JSBSim internals, or aircraft trim and have an idea why this happens or how to fix it — this is an open invitation to try. Ways to contribute:

• Open an issue or discussion on the PHI-CTRL GitHub repo with your findings, or a PR with a working fix and reproduction
• Use the Community tab on this dataset repo to post results, partial progress, or questions
• Fork this dataset, add a 7th attempt, and share what you found — clean success, partial success, or another documented failure are all useful

Whether the outcome is "here's the fix" or "here's attempt #7 and it also failed for reason X," either is a genuine contribution to the record.

---

Reproducing this

Each attempts/NN-*/run.py is self-contained (imports only from common/common.py) and can be run directly:

python attempts/01-grid-trim/run.py

Each run regenerates its own console_log.txt, telemetry_log.csv, ic_config.xml, and plot.png in its own folder. Run

python -c "import sys; sys.path.insert(0,'common'); from common import write_version_file; write_version_file('.')"

to regenerate jsbsim_version.txt for your own machine before publishing alongside these results, since JSBSim version drift could change the outcome and the dataset should always carry its own exact provenance.

Test conditions used

Altitude: 1000 ft, target airspeed: 90 kts — a modest, conservative cruise condition, not an edge case — held constant across all six attempts for a fair comparison. RUN_DURATION_S = 30.0 in common/common.py; increase this if you want to observe whether descent continues to actual gear contact rather than stopping at the -100 to -350 ft range recorded here.

Environment: JSBSim 1.3.1 (pip), Python 3.12.3, Linux. Exact c172p.xml provenance (path + SHA-256) is captured in jsbsim_version.txt in this repo — the hash, not the version number, is the authoritative identifier for which aircraft definition produced these results.

---

Citation

@dataset{bello2026c172pnegative,
  title  = {c172p Grounded: A Negative Result on JSBSim Airborne Trim Failure},
  author = {Bello, Mohammed Sani},
  year   = {2026},
  url    = {https://huggingface.co/datasets/SM-Bello/C172P-Grounded-JSBSim-Airborne-Trim-Failure-Negative-Result},
  doi    = {10.5281/zenodo.22218809},
  note   = {Related to PHI-CTRL: https://github.com/Sm-bello/PHI-CTRL}
}

---

Built by

Author	Mohammed Bello Sani
Lab	Penelope Inc. · PHI Lab
Institution	Air Force Institute of Technology (AFIT), Kaduna

Part of the PHI suite (PHI-Twin, PHI-Chain, PHI-CTRL, and related frameworks).