Towards Automated Safety Coverage and Testing for Autonomous Vehicles\n with Reinforcement Learning

Abstract

The kind of closed-loop verification likely to be required for autonomous\nvehicle (AV) safety testing is beyond the reach of traditional test\nmethodologies and discrete verification. Validation puts the autonomous vehicle\nsystem to the test in scenarios or situations that the system would likely\nencounter in everyday driving after its release. These scenarios can either be\ncontrolled directly in a physical (closed-course proving ground) or virtual\n(simulation of predefined scenarios) environment, or they can arise\nspontaneously during operation in the real world (open-road testing or\nsimulation of randomly generated scenarios).\n In AV testing, simulation serves primarily two purposes: to assist the\ndevelopment of a robust autonomous vehicle and to test and validate the AV\nbefore release. A challenge arises from the sheer number of scenario variations\nthat can be constructed from each of the above sources due to the high number\nof variables involved (most of which are continuous). Even with continuous\nvariables discretized, the possible number of combinations becomes practically\ninfeasible to test. To overcome this challenge we propose using reinforcement\nlearning (RL) to generate failure examples and unexpected traffic situations\nfor the AV software implementation. Although reinforcement learning algorithms\nhave achieved notable results in games and some robotic manipulations, this\ntechnique has not been widely scaled up to the more challenging real world\napplications like autonomous driving.\n