STEWART / LEARN
On this page

Evidence and limitations

A simulation is useful only when its claims fit its evidence. This project checks numerical identities, bounded responses and browser interactions. It has not been validated against a physical Stewart platform or an independent multibody engine.

What the checks cover

Layer Reproducible command What it establishes
Mechanics node tests/core.test.cjs IK/FK round trips, finite-difference Jacobians, mass/inertia properties, velocity-bias identity, joint efforts, controller responses and persistence checks on specified fixtures
Continuous scenarios node tests/endurance.test.cjs A 10 s moving-target case and 2 s passive, drives-off and all-active cases remain finite and respect tested bounds
Language/archive python tests/english.test.py Maintained source is English, document language is correct, decompressed Korean archive matches its fixed SHA-256
Browser python tests/browser_smoke.py Actual controls, joint changes, IK/FK, load/release, JSON download/import, CSV, geometry change, five lesson helpers, reduced motion and narrow-screen layout
Mode/recovery journeys python tests/usability.test.py Actual HTTP manual-to-ball return, reset/fall, preserved profile, FK/import invalidation, disabled controls, sensor inputs, keyboard/mobile and overlays
Ball/contact node tests/ball.test.cjs Named contact, control, sensor, fall and persistence cases, plus invalid inputs in ball_invalid.test.cjs
Generic servo/sensors node tests/realism.test.cjs Electrical/sensor/socket invariants; separate realism_dynamics.test.cjs balance, play, circle runs measure bounded trajectories
Teaching media python tools/record_demos.py Real UI actions and state-dependent observations generate captioned clips, with source hash and execution manifest
Website python tests/site.test.py Built entry points, local links, image alternatives, SVG descriptions, byte binding, caption files and 12 s video durations

Current implementation status belongs to the exact commit's Actions checks, not a permanently green label in this document. Generated raw results live in CI artifacts or local artifacts/ and tests/*-results.json.

English v1.1 preparation snapshot

The following is a dated local evidence snapshot, 2026-09-14, not a promise that every later commit passes. It used Node 22.16.0, Python 3.13.5 and Chromium 144.0.7559.96. The original mechanics and renderer were retained; the maintained UI was translated and lesson navigation added. The 26 core checks, four continuous scenarios and three language/archive checks passed. The browser interaction suite also passed.

Selected result Measured value Meaning
Core IK/FK maximum position error on tested poses about 5.06e-10 m Local numerical round trip, not a physical positioning accuracy
Core finite-difference body-Jacobian comparison about 1.14e-10 Derivative consistency for tested states
Core gravity-only hold over 0.5 s about 1.00e-11 m drift Gravity cancellation when plant and controller share the same model
Core IK tracking after 1.3 s about 0.149 mm position error One specified transient, not a universal tracking bound
Browser 16 N / 800 N/m, after 1.5 s settling 19.99947 mm X displacement Approximately the predicted 20 mm equilibrium
Browser after force release, another 1.5 s 0.0005004 mm X displacement Recovery in this model and fixture
Browser width/scroll width at mobile viewport 390 / 390 px No horizontal page overflow in that checked state
Browser JavaScript/console errors and app requests 0 / 0 / 0 For the local tested interactions, not a claim about hosting logs

The browser runner used page.set_content with the exact built standalone HTML because navigation was restricted in that environment. CPU Canvas rendering was exercised; GPU/WebGL rendering, real Pages navigation and arbitrary user devices were not established by that test. Browser timings can vary because ordinary UI interaction allows elapsed simulation time; deterministic core fixtures are the better numerical baseline. The recorded clips use explicitly stepped simulation time and must not be used to infer real-time throughput.

Mechanics source SHA-256 for this preparation snapshot:

f2d0375365e133fcc49b7542f393497d06f56aeb6b430c804e464ebccbf085f6

The generated build.json binds a particular English HTML file and the Korean original to source hashes. artifacts/recordings.json separately identifies the HTML used for recordings. These manifests should be inspected alongside the exact CI commit or downloaded package.

What a pass does not establish

Physics: no measured/calibrated actuator model, general collision mesh, backlash, structural flexibility, thermal duty model or experimental parameter identification. The optional generic servo does include averaged electrical response and drive loss; Ball Lab includes finite-deck frictional contact. See the maintained profile; the older preparation snapshot above is historical. Rendered fasteners and collars are not separately calibrated inertial parts. Ball-joint actuators are ideal three-axis torque sources, not a purchasable hardware design.

Kinematics: local FK does not enumerate all assembly modes or guarantee a unique pose. A conditioning threshold does not prove a globally reachable, collision-free or manufacturable workspace. Internal axial leg spin is not observable from six leg lengths.

Control: matching-model gravity cancellation does not prove robustness to unknown payload. A virtual spring/damper and bounded simulation do not prove passivity or stability for all gains, steps, configurations and active-joint sets. The bounded effort allocation is approximate, not a globally optimal constrained solver. Missing actuation cannot be repaired by a label saying “compliance”.

Numerics: mass/Jacobian/energy identities test implementation consistency, not independent physical truth. Damped semi-implicit stepping is not an exact integrator; penalty contacts can penetrate. The leg-orientation chart has an excluded singular direction. Energy and convergence studies require explicit trajectories, time steps, tolerances and acceptance criteria.

Product: keyboard-reachable controls, captions and numeric alternatives improve access, but there is no full accessibility certification or validated age-specific curriculum. The app makes no simulation-telemetry requests in the checked standalone run; the hosting provider can still keep ordinary access logs.

Before a stronger claim

For a research comparison, define the experiment, match conventions and inertial parameters in an independent engine, perform time-step refinement, and compare full trajectories, wrenches and energy with explicit tolerances. For physical hardware, independently verify mechanical limits, identification, sensing, timing and fail-safe control. This repository does not authorize or implement that connection.

See the model, equation-to-code map, experiment recipes and original references.