Cost Functions for Robot Motion Style

Abstract

We focus on autonomously generating robot motion for day to day physical tasks that is expressive of a certain style or emotion. Because we seek generalization across task instances and task types, we propose to capture style via cost functions that the robot can use to augment its nominal task cost and task constraints in a trajectory optimization process. We compare two approaches to representing such cost functions: a weighted linear combination of hand-designed features, and a neural network parameterization operating on raw trajectory input. For each cost type, we learn weights for each style from user feedback. We contrast these approaches to a nominal motion across different tasks and for different styles in a user study, and find that they both perform on par with each other, and significantly outperform the baseline. Each approach has its advantages: featurized costs require learning fewer parameters and can perform better on some styles, but neural network representations do not require expert knowledge to design features and could even learn more complex, nuanced costs than an expert can easily design.

Robot Expressive Motions

“…The optimization cost function can be formulated in different ways. First, it can be defined or learned to directly generate expressive motion (e.g., [104]) or, second, defined as a measure of how well people understand the robot, using an explicitly defined model of user inferences. For example, in Dragan and Srinivasa [17], the authors propose a mathematical formulation of a legibility functional, which is optimized to generate legible trajectories.…”

Human-Robot Interaction in a Shared Augmented Reality Workspace

“…From a robot's angle, it models humans by inferring goals [36], [37], tracking mental states [38], [39], predicting actions [40], and recognizing intention and attention [41], [42]. From a human agent's perspective, the robot needs to be more expressed [43], to promote human trust [44], to assist properly [45], [46], and to generate proper explanations of its behavior [44]. We believe the proposed shared AR workspace is an ideal platform for evaluating and benchmarking existing and new algorithms and models.…”

Teaching Robots to Span the Space of Functional Expressive Motion

“…Getting robots and virtual avatars to exhibit realistic looking and human-recognizable motions is a well-studied problem, from conveying intent in a task [4], [9], [10], to communicating incapability [11], [12], to expressing emotions [5], [13], [14]. In this section, we focus our attention on literature from the latter category, as our goal is enabling robots to learn emotive styles for performing functional tasks.…”

“…The motions in these methods are hand-crafted and, therefore, specific to the system and task they are being designed for. To generalize to a more diverse set of tasks, recent methods [5], [18], [19] try to learn a cost function that when optimized produces the desired emotive motion. However, these methods requires collecting labels for each emotion one at a time, resulting in inefficient and costly learning that fails to generalize to new emotions.…”

“…Taking an existing motion and adjusting its affective features, as researchers do with gestures [1], [4], would no longer meet the functional task specification. Instead, prior work [5] has proposed to learn a cost function *This research is supported by the Office of Naval Research (ONR-YIP), the Air Force Office of Scientific Research (AFOSR), the UCSF Weill Institute for Neuroscience, and the Apple AI/ML PhD Fellowship. All authors are with the EECS department at UC Berkeley {arjunsripathy, abobu, zhongyu li, koushils, dsbrown, anca}@berkeley.edu from user feedback for each desired emotion or style, that can be then optimized along with the task specification.…”