Drive GPT – An AI based generative driver model

Studies have shown that good driving alone can improve the energy efficiency of the vehicle by 30-40%, reduce accidents by 20% and reduce maintenance costs by 25%. Thus, providing feedback about driving behavior can help drivers optimize their driving habits enabling them to reduce the total cost of vehicle ownership while also increasing their safety. To solve this, we have built Drive GPT, an AI-based generative pre-trained transformer model, trained using 90 million data points from 100 vehicles driven across 8 cities within India over 6 months. The model has learned to generate good driving behavior under various road, traffic, weather, and vehicle conditions. Its architecture is based on a transformer network, which allows it to efficiently process sequences of driving steps. To make this a generic model and reduce bias, drives from across the cities and fleets have been oversampled and balanced to generate a representative driving behavior database on which the model is trained. Predictions on the test set comprising good driving behavior data resulted in R2 score of 0.98. Our results indicate that on average the drives generated by Drive GPT are 20% more energy efficient than bad drives. The Drive GPT model is used by Altigreen to estimate daily Drive scores for all the vehicles, provide personalized recommendations for drivers to improve their driving habits, and also give evaluation feedback during driver trainings. The article also compares and analyses key parameters such as acceleration, braking, energy efficiency, and regen concerning good and bad drives. Analysis also indicates considerable lesser use of throttle and braking in drives generated by Drive GPT, significantly enhancing component useful life