Much like humans in the real world, who observe other drivers and make inferences, we have designed a framework that treats human drivers as sensors to provide environment information to our intelligent system. We used probabilistic learning methods to estimate a sensor model that captures how people dynamically respond to pedestrians (i.e., learning the relationship between environment state and action), so that driver’s actions can then serve as a proxy for detection. This framework has shown significantly improvement in overall environment awareness.

Despite recent advances in learning and artificial intelligence, there remain many concerns in learning approaches to decision making and control that make them unreliable for safety critical systems. To address this, we have worked to improve learning algorithms to improve the safety and robustness of the resulting decision making and control policies.

Specifically, we have recently been considering sampling and training paradigms that will improve deep reinforcement and imitation learning policies. Through this work, we have made the following contributions: (1) we have considered a robust control take on deep RL, using adversarial reinforcement learning to improve model mismatch and safe transfer; and (2) we have improve interactive imitation learning by using a Bayesian approach that accounts for model uncertainty in the learning process and improves the quality of collected demonstrations obtained from human experts.

Before autonomous systems are unleashed into public domain, we must have a good understanding of where their vulnerabilities are and how likely they are to fail. However, many safety critical systems (like autonomous vehicles) suffer from very rare but high risk failures, meaning that it is very difficult to validate these systems by traditional means. To overcome this, we have been investigating an approach called adaptive stress testing to actively search for failure modes in simulation. This method uses reinforcement learning to control disturbances in your simulation to find the most likely failure. This has been applied to aircraft collision avoidance systems and to autonomous vehicles.

Behavior prediction remains one of the open problems for autonomous vehicles. Human driven vehicles (or generally the human-in-the-loop system) must be modeled in an accurate and precise manner that is easily integrated into control frameworks. Our developed driver modeling framework estimates the empirical reachable set, which is an alternative look at a classic control theoretic safety metric. This allows us to: (1) predict driving behavior over long time horizons with very high accuracy; (2) apply intervention schemes for semi-autonomous vehicles; and (3) mimic nuanced interactions between humans and autonomy in cooperative maneuvers.

To create autonomous systems that smoothly interact with human agents, the humans themselves must be carefully modeled. We have investigated a number of engineering approaches to design models and interactive systems, but also consider human factors, to ensure the systems reliably work in practice. The focuses of this work have considered: (1) human-in-the-loop test beds for data collection and empirical validation; (2) estimating and predicting discrete modes of behavior in vehicles; and (3) predicting human driver responses to conveyed intent.

By effectively designing autonomous systems that keep the user in mind, we can: improve driving performance while handing off control between the automation and the human driver; increase overall trust in the automation; and improve the predictability of the autonomous actions and overall situational awareness.