Personalizing Interfaces to Humans with User-Friendly Priors

TL;DR

This paper enables robots to adapt their interfaces to the current user so that the human's personalized interpretation is aligned with the robot's meaning, and biases the robot's interface towards these priors, resulting in signals that are adapted to the current user while still following social expectations.

Abstract

Robots often need to convey information to human users. For example, robots can leverage visual, auditory, and haptic interfaces to display their intent or express their internal state. In some scenarios there are socially agreed upon conventions for what these signals mean: e.g., a red light indicates an autonomous car is slowing down. But as robots develop new capabilities and seek to convey more complex data, the meaning behind their signals is not always mutually understood: one user might think a flashing light indicates the autonomous car is an aggressive driver, while another user might think the same signal means the autonomous car is defensive. In this paper we enable robots to adapt their interfaces to the current user so that the human's personalized interpretation is aligned with the robot's meaning. We start with an information theoretic end-to-end approach, which automatically tunes the interface policy to optimize the correlation between human and robot. But to ensure that this learning policy is intuitive -- and to accelerate how quickly the interface adapts to the human -- we recognize that humans have priors over how interfaces should function. For instance, humans expect interface signals to be proportional and convex. Our approach biases the robot's interface towards these priors, resulting in signals that are adapted to the current user while still following social expectations. Our simulations and user study results across $15$ participants suggest that these priors improve robot-to-human communication. See videos here: https://youtu.be/Re3OLg57hp8

Figures (3)

• Figure 1: A human is attempting to pass an autonomous car. The autonomous car communicates its policy (i.e., its driving style) using an LED mounted on the top of the car. Different users will interpret these LED signals in different ways. Although the interface does not know a priori how any specific human will interpret its signals, we recognize that there are underlying patterns all users expect interfaces to follow. Here the interface signals should recognize that there are two intuitive options that interpolate between (at one extreme) defensive driving and (at the other extreme) aggressive driving.
• Figure 2: Results from the Treasure (left) and Highway (right) simulations. Treasure: The human starts in state $s^{0}$ and tries to reach a hidden state $\theta$ that only the robot knows. The results shown here are averaged across $5$ simulations of $1000$ interactions in a $3$D environment. These results are consistent from $2$D to $8$D environments. Ours-C outperforms all other baselines in terms of average error ($p < 0.001$). Highway: The human (teal) tries to pass the autonomous vehicle (grey) without a collision. Here the results are averaged over $10$ simulations of $350$ interactions. Ours-C significantly outperforms baselines and Ours-P ($p < 0.001$). We note that the combination of Proportionality and Convexity are not shown: we found that these two priors conflict with one another during learning. Intuitively, this occurs because Convexity attempts to spread signals out, while Proportionality groups signals together.
• Figure 3: Left: Ours-C enables users to achieve a better performance in the Treasure (left) and Highway (right) tasks than LIMIT or Bayes. The difference in performance between methods is significant ($p < 0.0001$ (Treasure), $p \to 0.05$ (Highway)). An asterisk (*) denotes significance. Right: Users indicated that interfaces generated using the convexity prior (Ours-C) are more consistent and intuitive, and that they improved more over time with Ours-C.