Human-robot teams may be deployed in environments that require optimal task performance, but high stress environments can place considerable workload on the human, which can decrease task performance [4]. Task performance may be augmented by developing an adaptive teaming system that uses the human's workload state to determine how the robot interacts with the human. The human's overall workload state can be decomposed into ve components: cognitive, physical, auditory, visual, and speech [3]. Decomposing overall workload into components allows for an adaptive teaming system to determine if a robot interaction impacts the human, i.e., overload state, and how the interaction aects the workload state, i.e., will the cognitive component be overloaded. An adaptive teaming system can detect the human's workload state by using a workload assessment algorithm, which relies on objective workload metrics, e.g., heart-rate varability [2]. Workload assessment algorithms rely on machine-learning techniques to classify a Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for prot or commercial advantage and that copies bear this notice and the full citation on the rst page. Copyrights for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specic permission and/or a fee. Request permissions from permissions@acm.org. HRI'18 Companion, , March 5-8, 2018, Chicago, IL, USA © 2018 Association for Computing Machinery. ACM ISBN 978-1-4503-5615-2/18/03. . . $15.00 https://doi.org/10.1145/3173386.3176983 subset of the overall workload state; typically, only cognitive workload [2]. Detecting a subset of the overall workload state limits a workload assessment algorithm's viability for use in an adaptive teaming system, as the robot will be unable to determine how an interaction may impact the human. Thus, there is a need for the development of a workload assessment algorithm that can identify the overall workload state and that of each workload component.
The workload assessment algorithm assesses overall workload and each workload component every thirty seconds using knowledge of the task being completed. IMPRINT Pro workload models [1] of the tasks are used as the desired estimates during training.
The workload assessment algorithm relies on heart-rate (HR), heart-rate variability (HRV), skin-temperature (ST), respiration rate (RR), posture-sway (PS), and posture magnitude (PM) metrics of human workload. Each metric is ltered using adaptive exponential smoothing [5] and fed into a neural network. Each workload metric has a corresponding neural network that produces noisy estimates of a workload component, which are subjected to a weighted aggregation. The result is a workload component value, where the weights are determined by multivariate regression. The resulting equations for cognitive (W C ) and physical workload (W P ) are:
where w x represents the weightings for the corresponding metric x and b represents a bias term. The visual, speech, and auditory workload component values are determined by an IMPRINT Pro workload model. The overall workload value is the aggregate of the component values and is mapped to a workload state based on thresholds, which are the midpoint between the IMPRINT Pro workload models for each workload condition. For example, if 20 is the threshold value between low and high workload, then any value below 20 is classied as low workload.
Physiological data from two human-robot teaming evaluations, each focused on a specic interaction domain: peer and supervisory, were used to evaluate the algorithm. The workload models for each evaluation were generated using IMPRINT Pro.
Eighteen participants completed the peer-based evaluation, which required completing four fteen-minute rst response tasks with a Pioneer 3-DX robot. Each task elicited low or high workload.
The tasks required searching photographs for suspicious items, e.g., hazardous chemicals, on a Google Nexus 7 tablet; exhaustively searching an academic building hallway for hazardous or suspicious items; and sampling solid and liquid contaminants stored in containers based on instructions audibly given by the robot.
Thirty participants completed a supervisory evaluation, that required participants to monitor and control a simulated remotely piloted aircraft using the NASA Multi-Attribute Task Battery. Participants completed three fteen-minute trials corresponding to the underload, normal load, or overload workload states. Four concurrent tasks required participants to maintain a target in the middle of the cross-hairs; monitor two lights and four gauges and click on any out-of-range light or gauge; control fuel pumps to maintain the fuel levels of two tanks; and respond to air-trac control messages.
The results are presented for three trained workload assessment algorithms: trained solely on supervisory data (SUP), trained solely on peer data (PEER), and trained on equal proportions of both types of data (BOTH). Each trained algorithm was evaluated using the respective evaluation test data set, which was separate from the training data set. The average classication accuracies by evaluation and algorithm type are presented in Table 1. The algorithm achieves the highest accuracy when trained and tested on the corresponding data set, while classication accuracy decreases when the algorithm is trained on one data set and tested on the other. Specically, the PEER algorithm is unable to classify the overload state for the supervisory evaluation, which results in a low average classication accuracy. The SUP algorithm achieves a high cognitive and overall workload classication accuracy for the peer evaluation, but achieves a low physical workload classication accuracy. The BOTH algorithm performs exceptionally well for both evaluations, which demonstrates that the algorithm can be trained on multiple data sets with minimal impact on performance. 2. The highest correlations are achieved when the algorithm is trained and tested on data from the same evaluation. Additionally, each algorithm produced signicant correlations on data that the algorithm was not trained on, which illustrates that when workload changes; the algorithm's estimates respond accordingly, regardless of the algorithm's training data set. The presented diagnostic workload assessment algorithm is an essential step to realizing an adaptive teaming system. The classication results demonstrate that the workload assessment algorithm can accurately classify multiple cognitive, physical, and overall workload levels in dierent human-robot teaming paradigms. Accurately classifying overall workload and its components allows for an adaptive teaming system to identify the distinct components contributing to the overall workload state and potentially why the human is in the current workload state. Identifying the distinct workload components permits targeted adaptations to the components in order to normalize the workload state. An adaption based solely on the overall workload state cannot target a specic workload component; thus, the adaptation may be ineective.
The presented diagnostic workload assessment algorithm relies on objective workload metrics to accurately estimate overall workload and each workload component. Data collected from two human-robot teaming evaluations was used to evaluate the algorithm's capabilities, which showed that the algorithm accurately classied and tracked changes in cognitive, physical, and overall workload. The diagnostic workload assessment algorithm is an initial step to realizing an adaptive teaming system that is capable of determining a robot's interactions based on a human's workload level in order to augment task performance.