Automatic PID tuning for RoboCup SSL robots using artificial neural network, aiming for adaptive and robust control in dynamic competition environments.
This project consists of two main components:
ESP32-S3 firmware implementation for a three-wheeled omniwheel robot with adaptive PID control.
Key Features:
- Real-time PID control for three omniwheel motors arranged at 120° intervals
- Data collection and telemetry logging system
- Wireless communication with external inference station
- Motion pattern execution and trajectory tracking
- Client-server architecture for ML-based parameter optimization
Main Components:
main/- Core firmware implementation (control loops, movement calculation, telemetry)drivers/- Custom drivers (BLDC PWM, PID library, sensor fusion, WiFi)sensors/- AS5600 Hall effect encoder libraryplatform/- ESP32-S3 platform-specific code- Python servers for data logging and analysis
Technologies: ESP-IDF, FreeRTOS, C, Python
See pid-autotuning-firmware/README.md for detailed documentation.
Mathematical modeling, simulation, and machine learning environment for the omniwheel robot.
Key Features:
- Complete kinematic and dynamic model of three-wheeled omniwheel robot
- Non-linear system simulation with configurable parameters (wheel radius, gear reduction, robot dimensions)
- PID controller design, testing, and performance evaluation
- Trajectory planning and reference tracking simulation (linear, circular, figure-eight patterns)
- LSTM neural network training for automatic PID parameter optimization
- Training data generation from simulation trajectories
Main Components:
AutotuningPID_Simulator.ipynb- Robot dynamics, kinematics, and control system simulationAutotuningPID_RNNTraining.ipynb- LSTM model training and PID gain prediction- Data directories for training datasets and predictions
Technologies: Python, NumPy, SciPy, Matplotlib, Control Systems Library, TensorFlow/Keras, Scikit-learn, Jupyter
Model Architecture:
- LSTM layer (128 units, tanh) for sequence learning
- Dense layers for gain prediction
- Input: 15 features × 20 time steps (state, errors, control signals)
- Output: 9 PID gains (Kp, Ki, Kd for 3 wheels)
See Simulation/README.md for detailed documentation.
The autotuning system operates in a feedback loop:
- Firmware executes motion patterns and collects operational data
- Data is transmitted to the inference station
- LSTM Model (trained on simulation data) generates optimized PID parameters
- Parameters are sent back to the robot for improved control performance
This approach enables adaptive control that continuously improves based on real-world performance.
The LSTM model with 128 tanh units + 64 ReLU units demonstrated the best performance in simulation:
Performance Metrics:
- Oscillation Reduction: Significant decrease in tracking error variance
- Steady-State Error: Minimal offset after convergence
- Convergence Time: Faster settling compared to baseline PID
See Simulation/README.md for complete simulation results across all model architectures.
Real-world autotuning on ESP32-S3 hardware with 128 tanh units + 32 ReLU units:
Hardware Performance:
- Before Tuning: High oscillations and tracking errors
- After RNN Tuning: Smooth tracking with minimal overshoot
- Improvement: ~60-70% reduction in combined error score
Traditional tuning approaches tested on hardware:
Characteristics:
- Conservative gains
- Moderate oscillations
- Good stability margin
Characteristics:
- Faster response than Ziegler-Nichols
- Slightly higher overshoot
- Better for processes with dead time
Conclusion: RNN-based autotuning outperforms classical methods in both convergence speed and steady-state accuracy.
See pid-autotuning-firmware/README.md for complete experimental results across all model architectures and classical methods.