3. Timeline 2018–2026¶
Results relevant to learning-based control of liquid rocket engines, in date order. Engine rows are the main line; adjacent rows (landing guidance, thrusters, test-bench automation) are included where they shaped methods or show the state of the art next door. Every number links to where it appears; rows marked A rest on an abstract only (see §7).
{
"$schema": "https://vega.github.io/schema/vega-lite/v6.json",
"title": {
"text": "Every row of the table below, by date and kind of result",
"subtitle": "Hover a point for the source row. Dates given only as a year are placed mid-year"
},
"width": "container",
"height": 250,
"encoding": {
"x": {
"field": "date",
"type": "temporal",
"title": null,
"scale": {
"domain": [
"2018-06-01",
"2026-12-31"
]
},
"axis": {
"format": "%Y",
"tickCount": {
"interval": "year",
"step": 1
}
}
},
"y": {
"field": "lane",
"type": "nominal",
"sort": [
"Engine control, simulation",
"Hardware, engine or rig",
"Benchmark and infrastructure",
"Classical control",
"Adjacent fields"
],
"title": null,
"axis": {
"labelLimit": 220
}
}
},
"layer": [
{
"data": {
"values": [
{
"date": "2018-10-20",
"when": "2018-10",
"lane": "Adjacent fields",
"what": "Gaudet et al.: PPO, 6-DoF Mars landing",
"result": "landing error < 5 m; 4 % more fuel than optimum"
},
{
"date": "2019-06-01",
"when": "2019",
"lane": "Classical control",
"what": "Pérez-Roca et al.: LPRE control survey",
"result": "steady state mostly PID; transients open loop"
},
{
"date": "2019-07-10",
"when": "2019-07",
"lane": "Classical control",
"what": "Pérez-Roca et al.: MPC for transients",
"result": "MPC keeps pump-speed bounds that PID and LQR exceed"
},
{
"date": "2020-06-19",
"when": "2020-06",
"lane": "Engine control, simulation",
"what": "Waxenegger-Wilfing et al.: TD3 vs PID, gas-generator model",
"result": "IAE 519 (RL) vs 632 (PID)"
},
{
"date": "2020-10-12",
"when": "2020-10",
"lane": "Hardware, engine or rig",
"what": "Waxenegger-Wilfing et al.: hardware-in-the-loop concept",
"result": "≈ 250,000 flops per policy inference"
},
{
"date": "2021-03-17",
"when": "2021-03",
"lane": "Engine control, simulation",
"what": "Dresia et al.: SAC on the LUMEN model",
"result": "< 2 s load changes vs > 10 s open loop"
},
{
"date": "2021-06-01",
"when": "2021",
"lane": "Engine control, simulation",
"what": "Einicke: RL for pressure and mixture ratio (diploma)",
"result": "abstract only"
},
{
"date": "2021-08-01",
"when": "2021",
"lane": "Hardware, engine or rig",
"what": "Hörger et al.: robust RL, 22 N thruster",
"result": "basic functionality shown (abstract)"
},
{
"date": "2021-11-07",
"when": "2021-11",
"lane": "Adjacent fields",
"what": "Zou: rocket-recycling, open source",
"result": "landing success 14.6 % → 91.7 % by reward redesign"
},
{
"date": "2022-05-09",
"when": "2022-05",
"lane": "Adjacent fields",
"what": "Dresia et al.: FLAME test-bench valve sequences",
"result": "pump-inlet pressure within ±200 mbar"
},
{
"date": "2022-05-10",
"when": "2022-05",
"lane": "Hardware, engine or rig",
"what": "Hörger et al.: cold-gas RL test rig",
"result": "4 bar set point settles in 0.44 s"
},
{
"date": "2023-01-23",
"when": "2023-01",
"lane": "Hardware, engine or rig",
"what": "Dresia et al.: NN control of the LUMEN turbopump",
"result": "> 500 s of test, mean error < 1 %"
},
{
"date": "2024-03-02",
"when": "2024-03",
"lane": "Hardware, engine or rig",
"what": "Hörger et al.: zero-shot cold-gas controller",
"result": "RMSE ≤ 0.2 bar over 142 tests"
},
{
"date": "2024-03-15",
"when": "2024-03",
"lane": "Benchmark and infrastructure",
"what": "LUMEN put into operation at P8.3",
"result": "25 kN LOX/methane engine"
},
{
"date": "2024-07-21",
"when": "2024-07",
"lane": "Adjacent fields",
"what": "Jiang et al.: RAJS rocket landing",
"result": "success 8 % → 97 %"
},
{
"date": "2024-09-27",
"when": "2024-09",
"lane": "Adjacent fields",
"what": "Carradori: meta-RL, 6-DoF landing",
"result": "1000/1000 runs meet terminal constraints"
},
{
"date": "2024-10-14",
"when": "2024-10",
"lane": "Hardware, engine or rig",
"what": "Traudt et al.: first LUMEN hot-fire campaign",
"result": "4-valve closed loop at 20 Hz; 8 tests"
},
{
"date": "2024-11-01",
"when": "2024",
"lane": "Adjacent fields",
"what": "Urgolo et al.: learned fault monitors",
"result": "anomaly vs nominal F1 0.95"
},
{
"date": "2025-01-06",
"when": "2025-01",
"lane": "Hardware, engine or rig",
"what": "Hörger et al.: RL pressure and mixture ratio, 22 N thruster",
"result": "RMSE below 0.5 bar (abstract)"
},
{
"date": "2025-02-01",
"when": "2025-02",
"lane": "Adjacent fields",
"what": "Iafrate et al.: DRL first-stage landing",
"result": "metadata only"
},
{
"date": "2025-04-04",
"when": "2025-04",
"lane": "Benchmark and infrastructure",
"what": "Benchmark announced at RL4AA'25",
"result": "'freely accessible to the RL community'"
},
{
"date": "2025-04-09",
"when": "2025-04",
"lane": "Hardware, engine or rig",
"what": "Dresia thesis: SAC on the real engine",
"result": "1.3 % mean error over > 16 s"
},
{
"date": "2025-05-19",
"when": "2025-05",
"lane": "Benchmark and infrastructure",
"what": "Benchmark design: 2×2 task, 7 Gym test cases",
"result": "faults based on hot-run data"
},
{
"date": "2025-05-25",
"when": "2025-05",
"lane": "Engine control, simulation",
"what": "Bareiß: hybrid RL + PI (MSc)",
"result": "beats pure SAC and static PI (abstract)"
},
{
"date": "2025-06-30",
"when": "2025-07",
"lane": "Benchmark and infrastructure",
"what": "LUMEN model validated on hot fire",
"result": "MAPE 2.4 % on chamber pressure"
},
{
"date": "2025-07-02",
"when": "2025-07",
"lane": "Hardware, engine or rig",
"what": "Dauer et al.: out-of-capability detection",
"result": "retrained agent held pressure until test ended early"
},
{
"date": "2025-09-22",
"when": "2025-09",
"lane": "Benchmark and infrastructure",
"what": "DX'25 LUMEN diagnosis benchmark",
"result": "simulator on request only"
},
{
"date": "2026-03-31",
"when": "2026-03",
"lane": "Engine control, simulation",
"what": "Matanza et al.: curriculum PPO (poster)",
"result": "abstract only"
},
{
"date": "2026-06-11",
"when": "2026-06",
"lane": "Adjacent fields",
"what": "Chaudhary et al.: chance-constrained RL",
"result": "28 kg propellant overhead"
},
{
"date": "2026-07-29",
"when": "2026-07",
"lane": "Benchmark and infrastructure",
"what": "LUMEN simulator preprint",
"result": "errors below 5 % (abstract)"
},
{
"date": "2026-09-11",
"when": "2026-09-11",
"lane": "Benchmark and infrastructure",
"what": "DLR-RA GitHub organisation created",
"result": "empty on 2026-09-30"
},
{
"date": "2026-09-17",
"when": "2026-09-17",
"lane": "Benchmark and infrastructure",
"what": "LUMEN Control Challenge pitched",
"result": "not yet public"
}
]
},
"mark": {
"type": "point",
"filled": true,
"size": 90,
"color": "var(--viz-s1)",
"stroke": "var(--md-default-bg-color)",
"strokeWidth": 2
},
"encoding": {
"tooltip": [
{
"field": "when",
"title": "date"
},
{
"field": "what",
"title": "result"
},
{
"field": "result",
"title": "headline"
}
]
}
},
{
"data": {
"values": [
{
"date": "2025-04-09",
"lane": "Hardware, engine or rig",
"t": "1.3 % on the real engine"
},
{
"date": "2026-09-17",
"lane": "Benchmark and infrastructure",
"t": "Challenge pitched"
},
{
"date": "2024-03-15",
"lane": "Benchmark and infrastructure",
"t": "LUMEN in operation"
}
]
},
"mark": {
"type": "text",
"align": "right",
"dx": -8,
"dy": -11
},
"encoding": {
"text": {
"field": "t"
}
}
}
]
}| Date | Group | Method | Headline number | Link |
|---|---|---|---|---|
| 2018-10 | Univ. Arizona / MIT (Gaudet, Linares, Furfaro) — adjacent | PPO, 6-DoF Mars powered descent | Landing error < 5 m; 4 % more fuel than the 3-DoF optimal baseline | arXiv:1810.08719 |
| 2019 | ONERA / CNES / ArianeGroup (Pérez-Roca et al.) | Survey of LPRE control | Steady-state controllers "in most cases are of PID type"; transients "usually in open loop" (A) | doi:10.1016/j.paerosci.2019.03.002 |
| 2019-07 | ONERA / CNES / ArianeGroup (Pérez-Roca et al.) | Linear MPC + nonlinear pre-processor, Vulcain-like gas-generator engine, 5 valves | MPC respects turbopump speed bounds that PID and LQR exceed when throttling to 1.2× nominal \(p_{cc}\) | arXiv:1907.04273 |
| 2020-06 | DLR Lampoldshausen (Waxenegger-Wilfing, Dresia et al.) | TD3 (Stable-Baselines) on EcosimPro gas-generator engine start-up, 3 valves, 25 Hz | Episode reward −4.2 (RL) vs −6.5 (PID) vs −7.9 (open-loop sequence); IAE on \(p_{cc}\) 519 vs 632 vs 591 | arXiv:2006.11108, Table I; IEEE TAES 57(5), 2021 |
| 2020-10 | DLR (Waxenegger-Wilfing et al.) | Hardware-in-the-loop test concept for NN controllers | 400/300-neuron policy ≈ 250,000 flops per inference; a RAD750 (80 Mflop/s) runs it in real time | IAC-2020-C4.1.15 |
| 2021-03 | DLR (Dresia et al.) | SAC (RLlib) on LUMEN EcosimPro model, up to 6 valves, 10 Hz | Load changes 60→80 and 80→40 bar in < 2 s vs > 10 s open loop; constraint penalty 0 vs −162 | SP2020_533, p. 6–7 |
| 2021 | DLR / TU Dresden (Einicke) | RL for joint \(p_{cc}\) and \(R_{OF}\) control of an expander-bleed engine (A) | — | elib 142247 |
| 2021 (DOI year) | DLR (Hörger et al.) | Robust RL with domain randomisation, 22 N N₂O/C₂H₆ thruster | "preliminary experiments demonstrate the basic functionality" (A) | doi:10.2514/6.2021-3223 |
| 2021-11 | Zou — adjacent | PPO, 2-D rocket hover and landing (open source) | Landing success 14.6 % → 91.7 % from reward redesign alone, 2.1 M steps | rocket-recycling |
| 2022-05 | DLR + ESA FLAME (Dresia et al.) — adjacent | SAC finds tank-pressurisation valve sequences on the P5 test bench | LOX pump-inlet pressure held within ±200 mbar over a 700 s profile; ~1 day on 8 cores | SP2022_138 |
| 2022-05 | DLR (Hörger et al.) | SAC, cold-gas thrust chamber test infrastructure | 4 bar set point: settling 0.44 s, MSE 0.0055 bar² | SP2022_50, p. 7 |
| 2023-01 | DLR (Dresia et al.) | NN controller on the LUMEN oxidiser turbopump, P8.3 | > 500 s total test, mean error < 1 % | thesis p. 3; SciTech 2023 (A) |
| 2024-03 | DLR (Hörger et al.) | SAC, cold-gas chamber pressure, zero-shot sim-to-real, M11.5 | 142 hardware tests, RMSE ≤ 0.2 bar; trajectory RMSE 1.15 bar (hardware) vs 0.36 bar (sim) | Acta Astronautica 219, p. 9 |
| 2024-03 | DLR | LUMEN put into operation at P8.3 | 25 kN LOX/methane; first German methane engine | DLR LUMEN page |
| 2024-07 | Tsinghua / LandSpace (Jiang et al.) — adjacent | Random Annealing Jump Start + PPO/SAC family, 6-DoF + mass Simulink model | Landing success 8 % (PID guide) → 97 % | arXiv:2407.15083 |
| 2024-09 | TU Delft (Carradori) — adjacent | Meta-RL (LSTM, GTrXL) with PPO, 6-DoF atmospheric landing | 1000/1000 Monte-Carlo runs meet terminal constraints; 6 % more fuel than LQR-tracked optimum | MSc thesis |
| 2024-10 | DLR (Traudt et al.) | LUMEN hot-fire campaign; 4-valve closed-loop control at 20 Hz | 8 tests; ramp 40→60 bar within tolerance band | IAC 2024, p. 4–5 |
| 2024 | SAL / TU Graz / DLR (Urgolo et al.) — adjacent | Learned signal-temporal-logic monitors on a LUMEN-like EcosimPro model | Anomaly vs nominal F1 0.95, false-alarm rate 0.10 | DX 2024 |
| 2025-01 | DLR (Hörger et al.) | DRL \(p_{cc}\) + \(R_{OF}\) controller, 22 N N₂O/ethane thruster | RMSE "below 0.5 bar and 1"; "simulation is better than at the real system" (A) | SciTech 2025 |
| 2025-02 | Politecnico di Milano (Iafrate et al.) — adjacent | DRL first-stage propulsive landing (content not verified) | — (A, metadata only) | doi:10.1016/j.actaastro.2024.11.028 |
| 2025-04 | DLR / PLUS Salzburg (Dresia et al.) | Benchmark announced at RL4AA'25 | "will be made freely accessible to the RL community" | RL4AA'25 abstract |
| 2025-04 | DLR (Dresia, PhD RWTH) | SAC, 5 valves in sim vs MPC; 4 valves on hardware, 20 Hz, domain randomisation, no fine-tuning | Sim: 0.3 % / 0.1 % (\(p_{cc}\)/\(R_{OF}\)) vs MPC 1.4 % / 0.6 %. Hardware: 1.3 % mean over > 16 s | thesis p. 93, p. 113 |
| 2025-05 | DLR (Bareiß et al.) | Benchmark design: 2×2 MIMO, 7 Gym test cases | Seven test cases incl. domain adaptation and faults from hot-run data | AI4Aerospace abstract p. 55 |
| 2025-05 | DLR / RWTH (Bareiß, MSc) | Hybrid RL + PI on an electric-pump LOX/LNG engine model | Hybrid "learns a form of automatic gain scheduling", beats pure SAC and static PI (A) | elib 214433 |
| 2025-07 | DLR (Kurudzija et al.) | Validated LUMEN EcosimPro model on hot-fire data | MAPE 2.4 % (\(p_{cc}\)), 3.5 % (\(R_{OF}\)); 17 ignitions, > 20 min hot fire | EUCASS 2025-105 |
| 2025-07 | DLR (Dauer et al.) | Out-of-capability detection for the SAC engine controller | First hardware test oscillated (valve model); retrained agent held \(p_{cc}\) until the test was terminated early | EUCASS 2025-533 |
| 2025-09 | DX'25 (Pill et al.) | LUMEN fault-diagnosis benchmark (sibling to the control benchmark) | 15 sensor + 3 actuator + 3 component faults; simulator on request | DX'25 paper |
| 2026-03 | PLUS Salzburg / DLR (Matanza et al.) | Curriculum-guided PPO on the LUMEN environment (poster) | — (abstract only) | RL4AA'26 |
| 2026-06 | Univ. Auckland / ESA ACT / Univ. Bologna (Chaudhary et al.) — adjacent | Chance-constrained RL correction on an SCP trajectory, 2-D landing | 28.08 kg propellant overhead over the deterministic optimum, 100,000 samples | arXiv:2606.13605 |
| 2026-07 | DLR (Kurudzija et al.) | Transient system-level LUMEN-based simulator (preprint) | Errors "below 5 %" over the envelope (A) | doi:10.2139/ssrn.6806858 |
| 2026-09-11 | DLR Institute of Space Propulsion | GitHub organisation DLR-RA created |
1 repository (profile README) on 2026-09-30 | github.com/DLR-RA |
| 2026-09-17 | DLR at RL Bootcamp 2026 | LUMEN Control Challenge pitched: Gymnasium env, TFV/TOV actions, PPO and SAC demos | ~15 s hardware tracking before a sensor failure; public release in preparation | livestream 7:18:00 |
Affiliations are taken from the title pages of the papers I opened. The Kaiser 2021 Würzburg thesis on RL throttling for landing is omitted because no primary copy could be opened (§7.4).
2018–2022: methods in simulation
---
config:
theme: base
themeVariables:
fontFamily: Inter, sans-serif
cScale0: "rgba(57, 135, 229, 0.22)"
cScale1: "rgba(57, 135, 229, 0.34)"
cScale2: "rgba(57, 135, 229, 0.22)"
cScale3: "rgba(57, 135, 229, 0.34)"
cScale4: "rgba(57, 135, 229, 0.22)"
---
timeline
2018 : 6-DoF landing with PPO (Gaudet et al.)
2019 : LPRE control survey, PID dominates : MPC for transients (Pérez-Roca et al.)
2020 : TD3 beats PID on a simulated engine start-up (DLR) : HIL concept for NN controllers
2021 : SAC controls the LUMEN model with 6 valves (DLR) : rocket-recycling, 2-D landing
2022 : FLAME, SAC valve sequences for the P5 bench : cold-gas RL test rig2023–2026: hardware and the benchmark
---
config:
theme: base
themeVariables:
fontFamily: Inter, sans-serif
cScale0: "rgba(57, 135, 229, 0.22)"
cScale1: "rgba(57, 135, 229, 0.34)"
cScale2: "rgba(57, 135, 229, 0.22)"
cScale3: "rgba(57, 135, 229, 0.34)"
cScale4: "rgba(57, 135, 229, 0.22)"
---
timeline
2023 : NN turbopump control on hardware, below 1 % error
2024 : LUMEN in operation at P8.3 : zero-shot RL on a cold-gas rig : RAJS landing, 8 to 97 % success
2025 : Dresia thesis, 1.3 % on the real engine : benchmark announced and designed : model validated, MAPE 2.4 % : DX'25 LUMEN diagnosis benchmark
2026 : curriculum PPO poster : LUMEN simulator preprint : Control Challenge pitched, not yet publicReading the timeline¶
- Four years from simulation to hardware. DLR's first LUMEN RL controller ran in simulation in early 2021; the first closed-loop RL hot fire of the full engine is reported in the 2025 thesis. The intermediate steps were component-level hardware (turbopump 2023, cold-gas chamber 2024), which is where the valve-model and sensor-noise lessons were learned.
- Every hardware result is from one group. All engine-level RL hardware numbers come from DLR Lampoldshausen. The benchmark is the first attempt to let anyone else work on the same plant.
- The numbers are not on a common scale. Error metrics differ: mean absolute percentage error, IAE, RMSE in bar, summed reward. So do plants (gas-generator model, LUMEN model, LUMEN hardware, cold-gas rig). A shared evaluation service is exactly what the field lacks; §5 returns to this.