Planning for cars that coordinate with people: leveraging effects on human actions for planning and active information gathering over human internal state

Abstract

Traditionally, autonomous cars treat human-driven vehicles like moving obstacles. They predict their future trajectories and plan to stay out of their way. While physically safe, this results in defensive and opaque behaviors. In reality, an autonomous car's actions will actually affect what other cars will do in response, creating an opportunity for coordination. Our thesis is that we can leverage these responses to plan more efficient and communicative behaviors. We introduce a formulation of interaction with human-driven vehicles as an underactuated dynamical system, in which the robot's actions have consequences on the state of the autonomous car, but also on the human actions and thus the state of the human-driven car. We model these consequences by approximating the human's actions as (noisily) optimal with respect to some utility function. The robot uses the human actions as observations of her underlying utility function parameters. We first explore learning these parameters offline, and show that a robot planning in the resulting underactuated system is more efficient than when treating the person as a moving obstacle. We also show that the robot can target specific desired effects, like getting the person to switch lanes or to proceed first through an intersection. We then explore estimating these parameters online, and enable the robot to perform active information gathering: generating actions that purposefully probe the human in order to clarify their underlying utility parameters, like driving style or attention level. We show that this significantly outperforms passive estimation and improves efficiency. Planning in our model results in coordination behaviors: the robot inches forward at an intersection to see if can go through, or it reverses to make the other car proceed first. These behaviors result from the optimization, without relying on hand-coded signaling strategies. Our user studies support the utility of our model when interacting with real users. Keywords Planning for human-robot interaction • Mathematical models of human behavior • Autonomous driving This is one of several papers published in Autonomous Robots comprising the "Special Issue on Robotics Science and Systems".

Efficient and Trustworthy Social Navigation via Explicit and Implicit Robot–Human Communication

“…Legible motions are motions that express the robot's intent to observers, and can be different from predictable motions [26]. Sadigh et al modeled a human-robot system jointly as a dynamical system, and proposed methods to plan motions that actively influence human behaviors [28] and further actively gather information of human's internal state [29], [30]. Similar ideas were explored in human-robot collaboration scenarios.…”

“…In Eqs. (30) and (31), the reward at each time step is divided by show an example where initially the prediction doesn't match the measurement. We use the model to re-predict human actions at each time step, and the prediction starts to match measurement after two seconds.…”

Robots that Take Advantage of Human Trust

“…In order to identify communicative actions, we need a cognitive model of how the human learns: the human might respond to the robot's actions [8], infer the robot's objective [9], or follow a personalized learning strategy [10]. Our research builds upon these prior works [7]- [11] by identifying communicative actions for a trusting human model. Unlike research that explicitly generates deceptive actions [12], here the robot naturally selects misleading communication.…”

Defining interactions: A conceptual framework for understanding interactive behaviour in human and automated road traffic

“…Failure of a HAV to successfully interact with human road users may lead to congestion and human frustration, as a result of overly cautious behaviour on the part of the HAV (Millard-Ball 2016; Brown and Laurie 2017), or may even lead to crashes, if HAVs behave in ways that are unexpected by human road users (Alambeigi, McDonald, and Tankasala 2020). Improved understanding and models-both qualitative and quantitative-of how humans interact in traffic is a key prerequisite for vehicle manufacturers and software developers to program HAVs to successfully interact with humans (Camara et al 2019;Markkula et al 2018;Sadigh et al 2018;Schwarting et al 2019). This adds urgency to previously existing interaction-related research questions, and also introduces new research questions specific to human-machine interaction, for example whether and how eye contact or other human communicative gestures ought to be replaced with external human-machine interfaces Clamann 2015;Cefkin et al 2019).…”