Path tracking error control is an important functionality in the development of autonomous vehicles when a collision-free path has been planned. Large path tracking errors could lead to collision or even out of the control of the vehicle. Vehicle dynamic models are used to minimize the vehicle path tracking error so that control strategies can be designed under different scenarios. However, the vehicle dynamic model may not truly represent the actual vehicle dynamics. Furthermore, the nominal parameter employed in the vehicle dynamic model cannot represent actual operating conditions of the vehicle under environmental uncertainty. This paper presents a learning-based bias modeling method to improve the fidelity of any baseline vehicle dynamics model so that effective path tracking controller design can be achieved through a low fidelity but high-efficiency vehicle dynamic model with the aid of a few experiments or high fidelity simulations. The state-of-the-art of machine learning models, such as Gaussian process (GP) regression, recurrent neural network (RNN), and long short-term memory (LSTM) network, are employed for bias learning and comparison. A high-fidelity vehicle simulator, CARLA, is employed to collect virtual test data and demonstrate the effectiveness of the proposed bias-learning based control strategies under environmental uncertainty.