aactuator
EXPERIMENTS / FROM MICROSECONDS TO MINUTES

Follow the current.
Account for the heat.

Reproducible native experiments connect FOC, switch events, motor losses and thermal limits. Every curve states its model, units and evidence boundary.

01 / ASK WHAT THE MODEL ACTUALLY COMPUTES

Not a pose animation. Not a calibrated motor.

The live lab integrates current, rotor and output motion, DC-link voltage and lumped temperatures. These pages expose signals from the same native C++ components, rather than drawing ideal curves to look like test data. Here, “BLDC” means a brushless PM drive modeled as a sinusoidal PMSM; trapezoidal back-EMF and six-step commutation are not implemented. [1] [2]

Three distinct views: the browser runs an averaged inverter with 1-axis mechanics; the high-speed traces below run the native switched inverter; the MuJoCo films use a separate rigid-body/contact model. Matching geometry does not make their dynamics identical.

Loading source-bound native experiment data…

Numbers below come from the displayed dataset. Hover or tap a graph for a sample; line styles also identify series. Conditions and SHA-256 manifest · Full-resolution CSV + run configurations.

02 / FOC CONTROLS CURRENT, NOT POSITION DIRECTLY

Command, measurement and actual current are different signals.

A 4 A q-axis command is applied at 10 ms with both shafts locked. The PI controller changes voltage; resistance and inductance determine how current evolves. The ADC sees a filtered, quantized and delayed observation. The graph keeps each signal at its own timestamp, not a visually convenient shared timestamp.

A 4 A current step

Native averaged inverter; 50 μs control samples, linear analog-filter mode, seeded sensor noise. True current: interval endpoint. Reference and sensed dq current: control boundary. Not oscilloscope measurements.

Use the FOC coordinate calculator before interpreting q-axis current. The amplitude-invariant transform gives:

Pabc = 1.5(vd id + vq iq) Te = 1.5 p[ψm iq + (Ld − Lq) id iq] Pcu = R(ia² + ib² + ic²) = 1.5 R(id² + iq²)

For a balanced rotating sinusoid with id = 0, phase RMS is |iq|/√2. This does not mean each phase at locked rotor has that time RMS. DC-bus current is different again. Here R is per-phase, and flux is per-phase PM linkage; a line-to-line datasheet value cannot be pasted in without checking its convention. [2] [7]

03 / DUTY → SIX GATES → PHASE VOLTAGE

SVPWM is not the voltage waveform itself.

FOC supplies vd/vq. Inverse transforms produce three voltage references. This implementation adds the same common-mode term to all three, then normalizes by the DC bus. Equal offsets cancel in line-to-line voltage; they improve linear bus utilization. [3]

vcommon = −(max(va*, vb*, vc*) + min(va*, vb*, vc*))/2 dphase = 0.5 + (vphase* + vcommon)/Vdc |vdq| ≤ Vdc/√3 × (1 − 2m), m = 0.02 here

At 48 V, the configured vector limit is about 26.60 V, below the ideal 27.71 V because of the 2% duty margin. This is a voltage-vector bound, not phase RMS. Anti-windup tracks the final vector limit; it cannot guarantee stability under arbitrary delay. Controller limits · [4].

Three duty commands over one electrical revolution

Native svpwm() function sweep: 20 V rotating reference, 48 V bus, 2% margin. A modulation-function experiment, not a motor trajectory.

Look inside four actual PWM periods.

Choose a phase and a signal. Gates are boolean switch commands, not measured gate-source voltage. High and low gates are both off during dead time. Depending on current direction, a body diode can still conduct. Winding current changes continuously while terminal voltage switches. [5]

Switching and the winding response

Native switched inverter: 20 kHz PWM, 150 ns configured dead time, ≤1 μs integration split at switch events. Voltage/gates are held over the preceding interval; currents are interval endpoints. No interpolation across switching edges.

Never infer PWM ripple from the live lab’s 1 kHz plot: it displays an averaged inverter. This model includes Rds(on) drop and a fixed diode drop, but not MOSFET rise/fall waveforms, gate charge, switching-energy loss, reverse recovery, ringing or shoot-through. The illustrated PCB is not a validated gate-driver circuit. [6]

04 / GATES OFF ≠ CURRENT INSTANTLY ZERO

Stored energy still needs a path.

The next experiment injects a driver fault at 80 ms. The latch inhibits gates; the diode topology solver lets current decay rather than setting it to zero. At sufficient back-EMF, gate-off can also rectify energy into the bus. This locked-rotor example only demonstrates current extinction.

Current after a driver fault

Native gate-fault scenario; 4 A command, fault at 80 ms. Current samples at 50 μs. This is a supervisor-injected driver fault, not a comparator propagation-time measurement.

Independent FastTrip observes phase current outside the delayed ADC path. Comparator, blanking, BREAK, gate-off and supervisor acknowledgement have separate timestamps. Blanking can hide a real pulse; integration resolution still limits crossing-time evidence. STM32 automatic-output-enable options are configurable, so our explicit no-auto-rearm rule is a design choice, not a universal MCU guarantee. [10] [11]

05 / SECONDS OF MOTION, MINUTES OF HEATING

A cold motor and a warm motor are not the same plant.

Winding copper loss raises winding temperature; heat flows to the case and ambient. Resistance rises with temperature, increasing the voltage needed to hold current. A separate single RC state represents aggregate FET heating. [7]

R(Tw) = R25[1 + α(Tw − 25°C)], α = 0.00393/K Cw dTw/dt = Pcu − (Tw − Tc)/Rwc Cc dTc/dt = (Tw − Tc)/Rwc − (Tc − Ta)/Rca Cf dTf/dt = Pbridge − (Tf − Ta)/Rfa

Ten minutes at prescribed 12 A q-axis current

Prescribed-current thermal fixture, not a closed-loop drive. Both shafts locked; current forced to 12 A; 10 ms thermal steps, 1 s saved samples. The drive limiter is deliberately not part of this diagnostic.

Why distinguish this from an operating drive? A drive should reduce current before reaching its trip temperature. The existing thermal bench starts the winding at 105°C to test derating quickly; it is not evidence that a cold motor heats to 105°C in milliseconds.

A preheated drive limits the requested current

20 A requested after 10 ms, initial winding 105°C. Reference shown is the controller-limited value. Native Drive derates winding current linearly from 80°C to a 115°C trip, and FET current from 75°C to 100°C. These are synthetic configured limits, not recommended hardware ratings.

Only copper loss heats the winding model. Mechanical friction is removed as dissipated power, not routed into those thermal nodes. Iron/eddy losses, magnet-temperature effects, winding hotspots and PCB cooling geometry are omitted. Real continuous ratings require those effects and measured thermal parameters; this graph cannot establish a motor’s safe continuous torque. [8] [9]

06 / TORQUE IS NOT POWER; POWER IS NOT EFFICIENCY

Account for the watts before quoting a percentage.

The plant has separate motor/output inertias, a 6:1 gearbox, backlash and elastic coupling. Friction is smooth Coulomb plus viscous drag: Tf = bω + Tcoulomb tanh(ω/ε). This avoids a discontinuous sign but does not implement true static friction, stick-slip or Stribeck behavior. [12]

The example below commands 3 rad/s and adds a 1 N·m output load at 300 ms. The higher-friction case triples only output viscous drag from 0.02 to 0.06 N·m·s/rad; it is not a different motor product.

More drag requires more electrical input

Native closed-loop velocity experiments. Curves are 1 ms block means of 50 μs power diagnostics. The transient includes acceleration and stored energy; do not read every spike as a steady loss.
Pdc = Vdc Idc Pac = va ia + vb ib + vc ic Pdc − Pac = Pbridge (same evaluation state) Pac ≈ Pcu + Pfriction + Pgear + Pload + dEstored/dt ηmotoring = mean(Pload) / mean(Pdc), only in a stated quasi-steady motoring window

Boundary: inverter DC terminals → external load. Supply ESR, brake resistor and logic/driver auxiliary consumption are outside this ratio. The omitted iron and switching losses make it unsuitable as a commercial efficiency estimate. We report the remaining numerical energy residual instead of assigning it to an invented loss channel. At standstill, torque may be nonzero while Pload = Tload ωoutput = 0.

Regeneration reverses the power flow.

A velocity command drops from 5 to 0 rad/s at 300 ms. Negative inverter DC power means energy is returning to the link capacitor. The modeled source cannot sink current; a hysteretic brake resistor switches on at 53 V and off at 51 V. Those are example settings, not a protection design for a real board.

Braking can charge the DC link

Native averaged-inverter regeneration experiment. Separate units and axes; negative watts are not “negative efficiency”. No efficiency ratio is reported during regenerative or near-zero-power operation.
07 / HOW MUCH SHOULD YOU TRUST IT?

Trust a stated boundary, not a realistic-looking housing.

EffectImplementedStill outside the evidence
Motor and torqueSinusoidal dq R/L/flux, saliency torque, back-EMF, two inertias, compliance, backlashMagnetic saturation, cogging, harmonic back-EMF, manufacturing identification
Inverter and gatesAverage and switched models; dead time, Rds(on), diode topology, DC-link regenerationSwitching energy, parasitics, EMI, gate-charge dynamics, PCB validation
Sensing and protectionInline three-phase sensing, filter/ADC/offset/noise/delay, encoder quantization, faults and independent trip logic1/2 low-side shunt sampling-window reconstruction; measured STM32 ISR latency
Heat and efficiencyCopper R(T), winding/case/FET RC; modeled mechanical/bridge losses, signed powerIron and switching heat, hot spots, magnet weakening, measured efficiency map
ContactsBrowser 1-axis compliant stop; separate actual MuJoCo contactsGeneral browser 3D collision, hard zero-penetration guarantee, full coupling-rate convergence

Evidence comes in layers: analytical identities and rejection tests, full native regression, time-step studies, actual browser/MuJoCo execution, then physical calibration. The first layers are implemented; the last is NOT RUN. [13] [14]

Reproduce or change one assumption.

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build --parallel
./build/qdd_signal_lab results/my-signals
python tools/build_signal_data.py --exe build/qdd_signal_lab --out _site/data

Use a new evidence directory. Read the native experiment code, change one current/load/friction condition, and rerun. API guide owns integration details; annotated references own external reading. No copied controller is maintained for these plots.