Self-driving cars in mountainous terrain have many critical applications. For example, autonomous off-road vehicles for rescue and exploration, self-driving trucks for the mining industries, and autonomous shuttle services for remote resorts. Training autonomous vehicles in these environments has many challenges compared to urban or highway environments. These challenges include steep slopes, winding roads, and off-road obstacles. This research presents a new reinforcement learning model to drive in this complex environment. The model explores the effect of using fewer sensors for performance enhancement and cost-effectiveness in this environment. These sensors are RADAR, IMU, and GPS. Unity Engine with ML Agent is used as a simulator for training, as it provides a virtual environment for efficient and safe driving. The selected sensors facilitate a seamless knowledge transfer from the simulator to a real vehicle. They also provide sufficient information for measuring road gradient and angle between successive lane segments. A new reward system is designed based on the Proximal Policy Optimisation (PPO) algorithm with Behavioural Cloning (BC). The experimental findings suggest that the proposed model can navigate mountainous terrain safely and efficiently while avoiding collisions and staying in the correct lane. The model achieved a low total collision rate of approximately 0.04 collisions per obstacle, a lane exit rate of 0.014, maintained an average speed of 22 km/h, and a maximum speed of 48 km/h in complex simulated environments.