🤖 pid-ml-follower-model

Random Forest Regressor trained on 100 real robot runs to predict next-step line position error for a Hybrid PID + ML line following robot on Raspberry Pi.

📌 Model Overview

This model predicts next_error — the PID line position error at time t+1 — using 15 basic sensor inputs available at time t. The prediction is used as a residual correction on top of a classical PID controller, forming a Hybrid PID + ML control system.

Property	Value
Algorithm	`RandomForestRegressor` (scikit-learn)
Target	`next_error` = pid_error at t+1
Input features	15 sensor + PID state columns
Training rows	166,053
Test rows	42,930
Training runs	80 robot runs
Test runs	20 robot runs (zero leakage)

📊 Performance

Model	RMSE	MAE
Predict zero baseline	0.4252	0.3678
Persist current error baseline	0.0219	0.0124
Original RF (n=300, depth=15)	0.018756	—
This model (Optuna tuned)	0.018731	0.008295

14.3% improvement over persist-current-error baseline
Better than original RF on 13 / 20 test runs

⚙️ Hyperparameters

Parameter	Value
`n_estimators`	328
`max_depth`	22
`min_samples_leaf`	9
`max_features`	sqrt
Tuning	Optuna — 50 trials, TPE sampler

🔢 Input Features (exact order)

features = [
    "pid_error",        # current line position error
    "pid_error_prev",   # error from previous tick
    "pid_error_delta",  # error(t) - error(t-1)
    "pid_derivative",   # derivative term computed by PID
    "ir_centroid",      # weighted centre of IR readings
    "ir_spread",        # width of line across sensors
    "ir1_inv",          # IR sensor 1 (inverted)
    "ir2_inv",          # IR sensor 2 (inverted)
    "ir3_inv",          # IR sensor 3 (inverted)
    "ir4_inv",          # IR sensor 4 (inverted)
    "ir5_inv",          # IR sensor 5 (inverted)
    "left_speed",       # left motor speed
    "right_speed",      # right motor speed
    "speed_diff",       # left_speed - right_speed
    "loop_dt",          # time elapsed since last tick
]

Order is critical — must match exactly during inference.

🚀 How to Use

Install

pip install scikit-learn numpy huggingface_hub

Load Model

import pickle
from huggingface_hub import hf_hub_download

# Download model
model_path = hf_hub_download(
    repo_id  = "satwikshreshth1/pid-ml-follower-model",
    filename = "rf_model_tuned.pkl"
)

# Load
with open(model_path, "rb") as f:
    rf = pickle.load(f)

Inference

import numpy as np

features = [pid_error, pid_error_prev, pid_error_delta, pid_derivative,
            ir_centroid, ir_spread, ir1_inv, ir2_inv, ir3_inv, ir4_inv,
            ir5_inv, left_speed, right_speed, speed_diff, loop_dt]

predicted_next_error = rf.predict([features])[0]

Hybrid Control Loop on Raspberry Pi

# Alpha controls ML correction strength
alpha = 0.6    # recommended starting value
Kp    = 13.2   # must match your robot tuning

ml_correction = -predicted_next_error * Kp * alpha
ml_correction = np.clip(ml_correction, -10, 10)   # safety clip
final_output  = pid_output + ml_correction

Tuning Alpha

Alpha	Behaviour
0.0	Pure PID — ML disabled
0.3	Conservative hybrid
0.6	Standard hybrid (recommended)
1.0	Full ML correction

📦 Files in This Repository

File	Description
`rf_model_tuned.pkl`	Trained Random Forest model (pickle)
`model_meta.json`	Feature list, hyperparameters, metrics

🗂️ Training Data

100 real robot runs collected on 20 March 2026
221,967 total rows — 208,983 after cleaning
Robot: IR-based line follower on Raspberry Pi
PID tuning: Kp=13.2, Ki=0.0, Kd=0.475 | Base speed: 40
Loop frequency: ~65–70 Hz

Full dataset on Kaggle: satwikshreshth01/path-error-data

🔗 Related Links

GitHub Repository: satwik-shreshth/pid-ml-follower
Training Notebook: included in GitHub repo under notebook/
Related Project: Tripathagamini-S_Auto

📜 License

CC BY-NC 4.0 — Attribution required, non-commercial use only. https://creativecommons.org/licenses/by-nc/4.0/

👨‍💻 Author

Satwik Shreshth
MCA (Final Year), Sikkim University
satwikshreshth2002@gmail.com
@satwik-shreshth

Downloads last month: -