Robot Obstacles Avoidance System (Fuzzy System, Neural Networks)

An intelligent control system for autonomous robot navigation that decides speed and turning angle from the distance to the nearest obstacle and the angle to the target — implemented twice, as a rule-based fuzzy inference system and as an adaptive neuro-fuzzy inference system, so the two can be compared.

Obstacle avoidance mixes competing goals — keep moving toward the target, slow down near obstacles, turn away without oscillating — that are hard to express as crisp thresholds. Fuzzy control expresses them as overlapping linguistic rules; the question is how far a learned controller can reproduce and smooth that behaviour.

A Python research implementation in a Jupyter notebook built on scikit-fuzzy, NumPy, scikit-learn and matplotlib, with test cases for close, medium and far obstacles and extreme angles, and plots of membership functions, training error and FIS-versus-ANFIS output.

The fuzzy inference system maps distance (close, medium, far over 0–100 units) and angle (left, front, right over −90° to +90°) to speed (slow, medium, fast) and turn (left, straight, right) through 15 rules, triangular and Gaussian membership functions and centre-of-gravity defuzzification.

The ANFIS is a custom implementation with three membership functions per input and nine rules, trained by gradient descent through input, fuzzification, rule, normalisation and output layers, and evaluated with 5-fold cross-validation.

As reported in the repository, the ANFIS reproduces the fuzzy controller with a speed RMSE of about 3.35 (R² 0.98) and a turn RMSE of about 13.2 (R² 0.92); 5-fold cross-validation gives a speed MAPE of about 7.4 % and a turn MAPE of about 29 %. The work is educational and research-oriented, not a deployed robot.