Skill: rocketsmith · Paper: P. Pak, J. Barkley, R. Loghmani, D. Baich, A. Pamal, A. Barati Farimani, RocketSmith: Agentic Additive Manufacturing of High-Powered Rockets, arXiv:2606.00097 (2026) · Code: ppak10/RocketSmith
RocketSmith designed four FDM-printed Class 2 rockets with OpenRocket, build123d/CADSmith and PrusaSlicer behind an LLM agent. All four launched stably at a Tripoli Pittsburgh event. Two were recovered reflyable, and the altimeter apogees were 80–84% of the predictions.
The paper’s lesson, which RocketBench also found: let tools compute every number and have the LLM orchestrate.
Scope. Airframes for certified commercial motors within your certification level (H–I need L1, J–L need L2). Motor, propellant, igniter and ejection-charge design are out of scope; follow the manufacturer and your Range Safety Officer.
npx skills add BaratiLab/mechanical-engineering-skills --skill rocketsmith
pip install numpy matplotlib cadquery
cd skills/rocketsmith
examples/hp1_v38.json approximately reconstructs the paper’s High Power 1: 850 mm long, 101.6 mm
diameter, AeroTech H100W.
python scripts/rocket.py check examples/hp1_v38.json
| | This skill | Paper | |—|—|—| | Liftoff mass | 2268 g | 2249.8 g | | Stability, liftoff | 1.26 cal | 1.26 cal | | Stability, burnout | 1.38 cal | – | | Motor class | H (L1) | H (L1) |
All the checks pass.
How we got there: the first draft had 95 mm fins and stability 0.23 cal, an error. Moving the avionics forward and running
python scripts/rocket.py fins examples/hp1_v38.json --target 1.26 --out sized.json
widened the fin span to 162 mm, which is RocketSmith’s stability-analysis skill.
What the example is and isn’t. The fin span in hp1_v38.json is not taken from the paper: it
was sized with fins to reach the paper’s reported 1.26 cal. The stability match in the table is
therefore by construction. The flight comparison below, against the paper’s OpenRocket numbers, is
the check that was not tuned.
python scripts/rocket.py simulate examples/hp1_v38.json --cd 0.45 --plot hp1.png

| Apogee | Max velocity | Measured apogee | |
|---|---|---|---|
| HP1, this skill | 326 m | 75.0 m/s | 276 m |
| HP1, paper’s OpenRocket | 338.8 m | 76.4 m/s | |
| HP4 (J425R), this skill | 568 m | 113.7 m/s | 479 m |
| HP4, paper’s OpenRocket | 570.1 m | 115.2 m/s |
python scripts/rocket.py calibrate examples/hp1_v38.json --weighed "" --apogee 276 # Cd = 0.84
python scripts/rocket.py calibrate examples/hp4_big_heavy_2.json --weighed "" --apogee 479 # Cd = 0.76
The printed rockets flew with Cd ≈ 0.8, not the 0.45 that smooth airframes get. Printed surfaces, rail buttons, open slots and weathercocking all cost altitude. Use the fitted value for the next prediction.
python scripts/rocket.py cad examples/hp1_v38.json --out parts/

The output is an ogive nose cone with a shoulder, and a fin can with integrated fins and motor mount, as STEP and STL with mass estimates.
cadsmith-text-to-cad).calibrate --weighed nose=...,fincan=... to re-check stability with the
real masses.Cite: Pak et al., arXiv:2606.00097 (2026).