Blog · · 5 min read

Can AI run an FEA? What a stress check found in a jumping robot

A 30 cm jumping robot passed its walking load and failed the landing check. How an AI-run finite element analysis set up the load cases, what it found, and what a simulation still cannot tell you.

A robot that jumps also has to land. It sounds obvious, but sentences like this decide whether a design works or breaks on day one.

Jump & Run Robot is a public project on makemachine.ai: a small, low-cost humanoid of about 30 cm that people can build themselves. You drive it with an Xbox controller, and it should walk, run and jump. Before anyone prints a leg, one question needs an answer: will the legs survive what the robot does to them?

That is the job of a finite element analysis (FEA). This article follows the first FEA of the robot's legs, run by the simulation module of the project, and shows what AI does well in a stress check and where an engineer still has to decide.

What an FEA answers, and what it does not

An FEA splits a part into thousands of small elements, the mesh, and computes how each one deforms and how much stress it carries under the forces you define. The decision comes from a comparison: the highest relevant stress against what the material can take. That ratio is the safety factor. Above the target, the part passes. Below it, the part has to change.

What an FEA cannot tell you is whether you asked the right question. If the force is wrong, every color in the stress picture is wrong too.

Three load cases, written down first

The module starts by fixing the load cases and writing them down:

  • Service (normal use): twice the body weight, 39.7 N.
  • Landing: 300 N, the hit when the robot comes down from a jump.
  • Misuse: a sideways push of 100 N.

It checked three parts of the first CAD model, all meant to be printed in PETG: thigh, shank and foot. Printed plastic is weakest across its layers, so the check uses an allowable stress of 25 MPa for that direction. The safety factor is this allowable divided by the highest von Mises stress, with a target of 2 or more.

Writing assumptions down first turns a picture into a claim anyone can check.

A simulation you can rerun

The model comes from the CAD step, where AI writes the geometry as code instead of drawing it (see CAD as code). The simulation is a script as well: it splits the solid model into elements, applies the supports and loads, and solves the linear static problem.

The fragile step is choosing the faces for supports and loads. A guessed face number is a classic silent error, so the script selects faces by direction, position and size and prints the selection for checking.

The module's instructions also require an engineer's checks:

  • Equilibrium: the reaction forces must balance the applied load.
  • Plausibility: the part must bend the expected way, by an amount a hand calculation confirms.
  • Convergence: the part is solved on finer and finer meshes, here 2.0, 1.4 and 1.0 mm. If the governing stress still changes by more than 10 % between the two finest, the value is not settled.
  • Singularities: peaks right at a clamped edge grow with every refinement and mean nothing, so the rating uses the highest stress away from the supports.

The result: fine for walking, not for landing

This is what the first FEA found on 13 October. The numbers are safety factors; the target is 2 or more.

Part Service (39.7 N) Landing (300 N) Sideways (100 N)
Thigh 23 3.1 0.26
Shank 4.8 0.63 0.14
Foot 8.8 1.17 1.69

Normal use is no problem. Landing is another matter: the shank fails with 0.63, the foot falls short with 1.17, and only the thigh passes. A sideways push of 100 N fails for all three parts.

Equilibrium held in all nine cases. Convergence was mixed, and the report says so. The shank's landing stress changed by only 4 % between the two finest meshes. The foot's landing peak sits at a sharp inner corner and kept growing: 14.6, then 18.3, then 21.3 MPa. So the foot's 1.17 is a best case, not a precise value.

That is the best kind of bad news, because it arrived before anyone printed and dropped a robot. The report also names the fixes: a fillet of at least 3 mm at the shank's inner corner plus a thicker section there, and a fillet of at least 2 mm at the foot's corner. For the sideways push, the shank's stress would have to drop about 14-fold, which geometry alone is unlikely to deliver. The report suggests a mechanical stop, or accepting the case as misuse.

Why this FEA is already out of date

The design did not stand still. The newest concept version, "Hidden Joints", puts the knee motor flat in the top of the thigh, moves the ankle motor onto the knee axis and builds the legs from load-carrying printed shells. It estimates a mass of 3.22 kg and a landing peak of about 630 N.

Those are new parts and new loads, and no simulation has checked them yet. A simulation result expires the moment the geometry or the load changes. The honest status: the first legs failed the landing check, the new legs are untested.

What AI does well in FEA, and what it does not

Where it helps:

  • The tedious setup: meshing, face selection, solver input, nine cases, a convergence study.
  • Discipline: the same checks every time, written down.
  • Honesty about its own numbers: flagging a value that has not settled instead of rounding it into a verdict.
  • Turning a failure into a concrete place and size of change.

Where an engineer still has to look:

  • The loads. A landing is a short impact, and a linear static analysis replaces it with one equivalent force. Choosing that force is judgment.
  • The supports. Here the bores were fully fixed, a simplification the report lists itself. Real joints have play.
  • The material. A printed part's strength depends on printer, settings and print direction. Guide values are a start, not a datasheet.
  • The model's limits. Linear statics assumes small deformations and elastic material: no contact, yielding or fatigue.
  • The test. No simulation replaces printing a leg and dropping it.

Two kinds of simulation in one robot

The robot's balance controller learns in a physics simulation, where it practices jumping and landing before a real robot exists. The standing jump is done there; walking and running are still open. That simulation asks whether the robot stays upright. The FEA asks whether its parts survive doing so. A jumping robot needs both answers to be yes.

Follow the project

What the project needs next is a CAD model of the new legs and a fresh FEA with the new loads. You can follow every version, including the ones that failed, on the project page.

Watch hardware get built by AI. Follow public projects on makemachine.ai, read every result in full and get an email when there is something new.

Get started See the live projects

More from the blog

Open-hardware projects built with AI