Abstract
Proportional-Integral-Derivative (PID) controllers are fundamental to robotics, but manual tuning is time-consuming and requires expertise. This paper presents a machine learning approach that predicts optimal PID parameters from robot physical characteristics in milliseconds. Using a dataset of 1,000 robot configurations optimized via Nelder-Mead method, we trained a neural network to map robot parameters (mass, damping coefficient, inertia) to near-optimal PID gains. Our method achieves 78.1% average improvement over an adaptive baseline heuristic, 90.4% improvement over Cohen-Coon method, and 51.7% improvement over CHR method. Testing on 1,000 unseen robot configurations shows robust generalization with 100% success rate (all cases showing improvement). The approach reduces tuning time from seconds to milliseconds while maintaining superior performance. Statistical analysis confirms extremely significant results (Wilcoxon W = 0.00, p < 1e-10) with large effect sizes (Cohen's d = 5.66 vs baseline, 6.62 vs Cohen-Coon), indicating ML outperformed all baseline methods in all test cases. This work demonstrates that ML-based PID tuning can effectively assist manual methods for diverse robotic systems.