MADRaS : Multi Agent Driving Simulator

Autonomous driving has emerged as one of the most active areas research it promise making transportation safer and more efficient than ever before. Most real-world autonomous pipelines perform perception, motion planning action in a loop. In this work we present MADRaS, an open-source multi-agent simulator for use design evaluation algorithms driving. Given start goal state, task is to solve sequence position, orientation speed values order navigate between states while adhering safety constraints. These constraints often involve behaviors other agents environment. MADRaS provides platform constructing wide variety highway track scenarios where multiple can be trained tasks using reinforcement learning machine algorithms. built on TORCS, car-racing simulator. TORCS offers cars with different dynamic properties tracks geometries surface. inherits these functionalities from introduces support training, inter-vehicular communication, noisy observations, stochastic actions, custom traffic whose programmed simulate challenging conditions encountered real world. used create complexities tuned along eight axes well-defined steps. This makes particularly suited curriculum continual learning. lightweight convenient OpenAI Gym interface independent control each car. Apart primitive steering-acceleration-brake mode hierarchical track-position – that potentially achieve better generalization. uses UDP based client server model simulation engine agent. multiprocessing run agent parallel process efficiency integrates well popular libraries like RLLib. We show experiments single without