In this project we have built a simulation add-on to exhibit realistic traffic behavior and interactions for testing self-driving car algorithms. The cars inside the simulator are controlled such that they exhibit realistic traits as displayed by human drivers on roads. It is estimated that for government regulators to give fully autonomous vehicles the green light, at least 6 billion miles of testing and information-gathering will be required. This can be expensive and can lead to fatalities. Inadequate testing is also very dangerous as they fail to account for the nuances of human driving.

The behaviors are represented as rule based models and learning based models. These models are trained from data that we have collected from the real world. We have also built a data processing and parameter extraction pipeline to process the real world traffic camera data.

S.T.A.R.S is an attempt to decouple testing and data collection, thereby allowing self-driving cars to learn from an abundant pool of real-world data at a fraction of the cost and more importantly without putting human lives at stake. To accomplish this, we collect traffic data and build a model to learn realistic behavior as exhibited by human drivers from the data. The model is then used to control cars inside a simulator. We have a rule based model and a conditional imitation learning based model to achieve this.