Levitation Control Repository

This repository contains embedded firmware, sensor characterization tooling, and a physics-based simulation stack as well as control algorithm trials for the Texas Guadaloop maglev system.

This code is property of Texas Guadaloop, a student-led hyperloop team from the University of Texas at Austin.

Repository Structure

.
├── AltSensorTesting/       # Arduino firmware for Baumer inductive sensor characterization
├── sensor/                 # Python scripts to collect and fit sensor calibration data
├── loadCellCode/           # Arduino firmware for HX711 load-cell calibration
├── TwoCellMagChar/         # Two-cell magnetic characterization rig (validates Ansys data)
├── lev_sim/                # PyBullet simulation driven by Ansys sweep data
├── MAGLEV_DIGITALTWIN_PYTHON/  # Earlier single-axis analytical digital twin (reference)
├── serial_plotter.py       # Live serial plotter utility
├── equilibrium.py          # Equilibrium gap/current solver
└── requirements.txt        # Python dependencies

Embedded Stack

`AltSensorTesting/`

Arduino firmware for reading Baumer inductive analog distance sensors. Uses interrupt-driven ADC sampling at ~77 kHz (16 MHz / prescaler 16) for low-latency gap measurement. Tracks the 10 lowest and 10 highest in-range ADC values over a sampling window to help establish calibration bounds. An out-of-range (OOR) digital pin is monitored in the ISR to discard invalid readings.

Key details:

ADC ISR at ~77 kHz; readings discarded automatically when OOR pin is HIGH
Serial commands: 1 to start sampling, 0 to stop and print boundary statistics
Baud rate: 2,000,000

Sensor Characterization

`sensor/`

Python pipeline for converting raw ADC readings from the inductive gap sensors into calibrated millimeter distances.

sensorCollector.py — Serial interface that reads live ADC values from Arduino and applies the calibration model in real time.
analogFitter-*.py — Curve-fitting scripts (polynomial, exponential, 3/4/5-parameter logistic) that fit calibration sweep data (data*.csv) to find the best sensor model. The 5-parameter generalized logistic form was found to give the best fit.
Sensor*Averages.csv / data*.csv — Raw and averaged calibration data for sensors 0–3. Sensor 3 required a different voltage divider (20 kΩ / 50 kΩ) because the induction sensor output exceeds 6 V.

Calibration constants (A, K, B, C, v) for the generalized logistic model are embedded directly in sensorCollector.py and sensor_simplerNew.py for deployment.

Magnetic Characterization (`TwoCellMagChar/`)

A minimal bench rig used to validate Ansys Maxwell FEA force predictions experimentally.

TwoCellMagChar.ino — Drives two H-bridge coil channels across a sweep of PWM values (−250 to +250 in steps of 50) while reading two HX711 load cells simultaneously. Averages 10 measurements per PWM step and reports gram-force readings over serial.
MagCharTrial.xlsx — Recorded force-vs-PWM data from physical trials.
CalibConsts.hpp — Load-cell offset and scale constants shared with loadCellCode/.

The measured force curves confirmed that the Ansys sweep data is in reasonable agreement with physical hardware, providing confidence for using the FEA model inside the simulation.

Levitation Simulation (`lev_sim/`)

A PyBullet-based simulation environment for 4-point active magnetic levitation of the pod. The simulation is driven by an Ansys Maxwell parametric sweep (coil currents × gap height × roll angle → force & torque), fitted to a polynomial regression model for fast inference.

Data pipeline

Ansys sweep — Ansys Results 12-9.csv / .xlsx contains FEA results sweeping left/right coil currents, roll angle, and gap height.
Function Fitting — Function Fitting.ipynb fits the Ansys data to a PolynomialFeatures + LinearRegression model (inputs: currL, currR, roll, 1/gap). The trained model is saved to maglev_model.pkl.
Fast inference — maglev_predictor.py (MaglevPredictor) loads the pickle and bypasses sklearn overhead by extracting raw weight matrices, running pure-NumPy polynomial expansion for ~100× faster per-sample prediction.

Simulation environment

lev_pod_env.py implements a Gymnasium Env wrapping PyBullet:

State: 4 gap heights (normalized) + 4 gap-height velocities
Action: 4 PWM duty cycles ∈ [−1, 1] (front-left, front-right, back-left, back-right coils)
Physics: first-order RL circuit model per coil (mag_lev_coil.py); coil parameters: R = 1.1 Ω, L = 2.5 mH, V_supply = 12 V, I_max = 10.2 A
Forces: MaglevPredictor.predict() maps coil currents + gap + roll → force and torque applied to the pod body in PyBullet
Noise: configurable Gaussian sensor noise (default σ = 0.1 mm) and stochastic disturbance forces
Target equilibrium gap: 11.86 mm (9.4 kg pod)
Simulation timestep: 1/240 s

The environment is controller-agnostic — any control algorithm can be plugged in:

File	Controller
`pid_simulation.py`	Feedforward LUT + PID
`lev_PID.ipynb`	Interactive PID tuning notebook
`optuna_pid_tune.py`	Optuna-based automated PID hyperparameter search
`lev_PPO.ipynb`	PPO reinforcement learning via Stable-Baselines3

Pre-tuned PID gain sets are saved in pid_best_params*.json (variants for 1500, 3000, and 6000 Optuna trials).

PWM circuit model

PWM_Circuit_Model.py is a standalone electrical model of the H-bridge PWM drive circuit used for offline verification of the coil current dynamics.

Setup

pip install -r requirements.txt

Key dependencies: pybullet, gymnasium, stable-baselines3, scikit-learn, optuna, pyserial, numpy, scipy, pandas, matplotlib.

5.9 KiB Raw Blame History Unescape Escape