Q-Learning Robot: Reinforcement Learning Simulator

This simulator trains a robot agent to cross a grid arena using real tabular Q-learning: the agent maintains one Q-value per (state, action) pair, acts epsilon-greedily, and after every step updates its estimate with Q(s,a) += α·(reward + γ·max(Q(s′,a′)) − Q(s,a)). Episodes run continuously — the robot starts over each time it reaches the goal, falls in a pit, or runs out of steps — while the grid's color heatmap and policy arrows reveal exactly what the agent has learned so far. Tune the learning rate, discount factor and exploration decay to see how each reshapes the speed and stability of convergence.