ML-Based PID Auto-Tuning
Paper
Machine Learning PID Parameter Optimization
Oct 2025 - Nov 2025
The Problem
Manual PID tuning takes 2+ hours per robot and requires expert knowledge. Classical methods (Cohen-Coon, CHR) are suboptimal.
The Solution
ML model (MLP: 3→128→64→32→3) trained on 1,000 robot configurations optimized via Nelder-Mead method. Predicts Kp, Ki, Kd from mass, damping coefficient, and inertia in <0.2ms.
Key Innovation:
- •<0.2ms inference time
- •100% success rate across 1,000 test cases
- •Statistical validation (p<1e-10, Cohen's d=5.66-6.62)
- •Noise robustness (tested 0-20% noise)
- •Dimensionally correct physics model
- •Comprehensive baseline comparisons
Results
78.1% improvement vs adaptive baseline, 90.4% vs Cohen-Coon, 51.7% vs CHR. 100% success rate (1,000/1,000 test cases). Prediction time: 0.179±0.013ms
└─ Visualization results: https://github.com/ilyasidk/ml-pid-optimization/tree/main/results
└─ Improvement distribution, noise robustness, and comparison charts available
└─ Statistical analysis with bootstrap CI and multiple effect sizes
Tech Stack
Backend
- ├─Python
AI & Integrations
- ├─Neural Networks (MLP)
- ├─TensorFlow/PyTorch
- ├─Nelder-Mead Optimization
Infrastructure
- ├─Custom Robot Simulation
- ├─Statistical Analysis
- ├─Matplotlib Visualization