Carlos Toledo · instruments · the physics answer key

Every number a simulator writes is an exact fraction.

So whether a recorded run kept its energy, to any tolerance you name, can be decided with whole numbers and no rounding. Three scenes whose physics is known exactly, run in two engines built for simulating robots: MuJoCo 4.0.1 with each of its five integrators, and Newton 1.8.0 with two of its solvers, at their defaults and with the angular damping off. Pick a run: your tab fetches the recorded bytes and decides them again.

engine
solver
integrator
timestep, seconds

What the 100 recorded runs say.

In each cell: the largest distance from the law anywhere in the run, rounded up, and the tightest of six tolerances the whole run stays inside. Configurations that wrote the same bytes share a row. Nothing was refused: the 1,464 verdicts are 393 Certified and 1,071 Refuted.

The tumbling body: spin energy against its starting value, over 10 seconds

Newton 1.8.0, solvertimestep 0.01 stimestep 0.005 stimestep 0.002 stimestep 0.001 s
SolverSemiImplicit, defaults60.6%−60.6% at the endholds at no tolerance tried61.9%−61.9% at the endholds at no tolerance tried62.7%−62.7% at the endholds at no tolerance tried62.9%−62.9% at the endholds at no tolerance tried
SolverSemiImplicit, damping off9.94%+9.94% at the endholds at no tolerance tried4.83%+4.83% at the endholds at no tolerance tried2.00%+2.00% at the endholds at no tolerance tried1.16%+1.16% at the endholds at no tolerance tried
SolverFeatherstone, defaults; SolverFeatherstone, damping off9.93%+9.93% at the endholds at no tolerance tried4.79%+4.79% at the endholds at no tolerance tried1.88%+1.88% at the endholds at no tolerance tried0.930%+0.930% at the endholds to 1%

Newton’s SolverSemiImplicit, built with every argument at its default, loses 60.6% to 62.9% of the spin energy in 10 seconds with no force acting, and the energy never rises at any step. Its default includes an angular damping of 0.05; with that argument set to zero the same solver gains energy instead (+2.00% at 0.002 s). SolverFeatherstone wrote the same bytes with the default and with zero, in every scene at every timestep.

MuJoCo 4.0.1, integratortimestep 0.01 stimestep 0.005 stimestep 0.002 s the defaulttimestep 0.001 s
Euler9.93%+9.93% at the endholds at no tolerance tried4.79%+4.79% at the endholds at no tolerance tried1.88%+1.88% at the endholds at no tolerance tried0.931%+0.931% at the endholds to 1%
implicitfast, discrete8.65%−8.65% at the endholds at no tolerance tried4.47%−4.47% at the endholds at no tolerance tried1.83%−1.83% at the endholds at no tolerance tried0.919%−0.919% at the endholds to 1%
implicit8.65%−8.65% at the endholds at no tolerance tried4.47%−4.47% at the endholds at no tolerance tried1.83%−1.83% at the endholds at no tolerance tried0.919%−0.919% at the endholds to 1%
RK41.91×10−9−4.55×10−10 at the endholds to 1 in 1061.06×10−10−1.48×10−11 at the endholds to 1 in 1092.51×10−12−1.67×10−13 at the endholds to 1 in 1091.62×10−13−1.47×10−14 at the endholds to 1 in 1012

With MuJoCo’s Euler, which is still that engine’s default, the spin energy rises at the very first step and is +1.88% after 10 seconds at the default timestep. With implicitfast, the integrator its documentation recommends, it is −1.83%: the energy never rises at any step of any of those runs, exactly, and it leaks instead. With RK4 the largest change is 2.51×10−12.

The falling body: height against the formula, over 2 seconds

MuJoCo 4.0.1, integratortimestep 0.01 stimestep 0.005 stimestep 0.002 s the defaulttimestep 0.001 s
Euler, implicitfast, implicit, discrete98.1 mmholds at no tolerance tried49.1 mmholds at no tolerance tried19.7 mmholds at no tolerance tried9.82 mmholds to 1 cm
RK42.35×10−14 mholds to 1 pm3.59×10−13 mholds to 1 pm1.67×10−13 mholds to 1 pm2.09×10−12 mholds to 1 nm
Newton 1.8.0, solvertimestep 0.01 stimestep 0.005 stimestep 0.002 stimestep 0.001 s
all four configurations (the same bytes)98.1 mmholds at no tolerance tried49.1 mmholds at no tolerance tried19.7 mmholds at no tolerance tried10.1 mmholds at no tolerance tried

In MuJoCo at the default timestep the body ends 19.7 mm from where the formula puts it, and passes 1 cm of error at step 510. Its speed is right to the last digits; only the position lags, which is what a first-order step does. With RK4 the position stays within a millionth of a millionth of a metre at the same timestep. Newton’s position lags the same way (first past 1 cm at step 511 at 0.002 s); its speed holds to a millimetre per second and not to a micrometre per second, in 32-bit floats.

The pendulum: energy against its starting value, over 10 seconds

MuJoCo 4.0.1, integratortimestep 0.01 stimestep 0.005 stimestep 0.002 s the defaulttimestep 0.001 s
Euler, implicitfast, discrete1.85%holds at no tolerance tried0.920%holds to 1%0.367%holds to 1%0.184%holds to 1%
implicit1.85%holds at no tolerance tried0.920%holds to 1%0.367%holds to 1%0.184%holds to 1%
RK41.52×10−8holds to 1 in 1066.68×10−10holds to 1 in 1091.29×10−11holds to 1 in 1097.18×10−13holds to 1 in 1012
Newton 1.8.0, solvertimestep 0.01 stimestep 0.005 stimestep 0.002 stimestep 0.001 s
SolverFeatherstone, defaults; SolverFeatherstone, damping off1.85%holds at no tolerance tried0.920%holds to 1%0.367%holds to 1%0.184%holds to 1%

Two published sentences, decided.

None of this is a fault report. An integrator of finite order is expected to move away from the true motion, and MuJoCo’s documentation describes the behaviour above. What this page adds is the decision: each statement checked on the recorded bytes with whole numbers, the step where it first fails, and the same check run again in your tab.

One sentence of that documentation is a claim exact enough to decide: that under its implicit integrators the spin energy of a free body never increases. On the 8 recorded runs it covers, it holds at every step with no tolerance at all. (The four discrete runs are the same bytes as implicitfast in this scene, so it holds there too.)

The kinetic energy of a spinning free body is non-increasing in the absence of applied force. Steady spins about principal axes are conserved almost exactly; tumbling motion is mildly damped […]. Systems requiring long-horizon energy conservation of tumbling bodies should use the RK4 integrator.
MuJoCo documentation, Computation, Numerical integration, “Gyroscopic derivatives for free bodies”, read 17 November 2026. Source.

Newton’s class description says its semi-implicit method preserves energy. That is a description of the method, with no bound. The solver’s constructor also takes an angular damping, documented, with the default 0.05, and a body built with the defaults is damped by design: on the 4 recorded runs the spin energy is more than 1% down within 0.11 seconds and 60.6% to 62.9% down after 10.

That is a documented default beside a sentence, not a bug: passing angular_damping=0.0 turns it off, and Newton’s own verification tests do. SolverFeatherstone carries the same sentence and the same default, and wrote identical files with and without it. The report has the detail.

Semi-implicit time integration is a variational integrator that preserves energy, however it not unconditionally stable, and requires a time-step small enough to support the required stiffness and damping forces.
Newton 1.8.0, the class description of SolverSemiImplicit and, in the same words, of SolverFeatherstone, quoted as written. Source.

How a float gets decided.

The bytes. A run is a file: the engine’s time, positions and velocities after every step, as it wrote them. Each of the 100 files is named by its sha256, and your tab checks that before deciding anything.

The fraction. A float, 64-bit or 32-bit, is a whole number times a power of two. Energy and angular momentum are sums of products of those, so they are fractions too, and “within 1%” becomes a comparison of two whole numbers. Only the pendulum needs a cosine, which is trapped between two fractions as tightly as the comparison needs.

Twice. The record was written only after two programs that share no code, one in JavaScript and one in Python, agreed on every field of every run, and after 12 planted files with a known answer were each decided as planted.

The rules were written and committed before the first simulation step of each engine; one law was added after the MuJoCo runs and is marked as such. The report has the rules, the limits and what is still open.

contact

Write to Carlos

Sent by Web3Forms to Carlos’s inbox; nothing stored here.

Or write to carlos@carlostoledo.co, or open the public pilot form on GitHub.