Adaptable (Not Adaptive) Automation: Forefront of Human–Automation Teaming

Introduction

The increasing examination of automation, autonomous agents, and intelligent systems in a variety of application domains is evident in Human Factors and related publications. For example, advancements in hardware/software to automatically support tasks normally ascribed to humans have and are more frequently being considered in aerospace systems, surface transportation, health care, occupational ergonomics, and robotics (Davenport & Kalakota, 2019; Moray et al., 2000; Sheridan, 2016; Young et al., 2007).

The key goal of this article is to highlight a research gap pertaining to how automation is applied in systems. After a mini-review of flexible design approaches for implementing human–automation systems, the tradeoffs pertaining to the locus of authority that determines how levels of automation are assigned to tasks will be detailed. This will be followed by suggestions for research to maximize the benefits of flexible application of automation for more effective human–automation teaming. For this article, the term “automation” is used for any system that accomplishes (partially or fully) a function/task that was previously or conceivably could be completed by a human operator (Parasuraman et al., 2000), including “autonomy” developments that are designed to be able to achieve task goals independently without external interventions.

Design Approaches for Human–Automation Systems

In some cases, system applications of automation provide vital safety measures, such as taking over control of a high-speed vehicle when it is detected that the human has lost consciousness or when a control input (e.g., to avoid a collision) is needed in less time than is humanly possible. In such instances, a fixed application of automation is appropriate—whenever the critical condition is identified, the safety measure automatically activates. Aside from life-threatening situations, research conducted to date has generally demonstrated that fixed applications of automation to alleviate workload can result in reduced mode and situation awareness, mistrust, complacency, and skill degradation (Endsley & Kaber, 1999; Endsley & Kiris, 1995; Kaber & Endsley, 2004; Parasuraman & Riley, 1997; Parasuraman et al., 1996).

In contrast, design approaches that flexibly apply automation such that its level and/or type changes during system operations can help humans avoid out-of-the-loop problems and have been shown to enhance human–automation system performance (Inagaki, 2003; Kaber & Endsley, 2004; Parasuraman et al., 1996; Scerbo, 2007). In this manner, the system is designed to accommodate a granularity of control (Endsley, 2017), ranging from low-level automation that involves the human more (selecting courses of action and making inputs to determine how the system functions) to high-level automation that engages the human less—the automation determines the best courses of action and how the system functions. At lower level(s) of automation (LOAs), the human has more opportunity to interact with the system, providing a better basis for building trust in its operation and maintaining situation awareness and skills. However, lower LOAs can increase the human’s workload, as well as stress and/or fatigue. At higher LOAs, the human’s workload is less, but this comes with potential costs to situation awareness, engagement, and skill maintenance, as well as other problems if the system’s automated changes are not timely or context relevant.

Applications, however, are not limited to low and high LOAs. Each individual task involved in system operation can be assigned a different LOA. More importantly, a range of LOAs can be implemented in the system including multiple intermediate levels defined in a continuum with fully manual and fully automated LOAs at the two extreme ends. Each intermediate level differs in its specification of the degree to which a task is automated. Perhaps the earliest and most widely cited taxonomy is from Sheridan and Verplank (1978), which defines 10 levels ranging from fully manual to fully automated (Table 1), with each LOA defining the respective roles of the human and automation and how they are to cooperate with each other. This taxonomy has been extended in many ways (see review by Vagia et al., 2016), including multidimensional ones that consider multiple factors (e.g., Parasuraman et al., 2000). See also the “Special Issue on Advancing Models of Human–Automation Interaction” for a discussion of the construct of LOAs (Roth & Pritchett, 2018).

OPEN IN VIEWER

Table 1

Levels of Automation (Based on a Scale Proposed by Sheridan & Verplank, 1978 )

	Levels of Automation of Decision and Action Selection
High	10. The computer decides everything, acts autonomously, ignoring the human.
	9. informs the human only if it, the computer decides to
	8. informs the human only if asked, or
	7. executes automatically, then necessarily informs the human, and
	6. allows the human a restrictive time to veto before automatic execution, or
	5. executes that suggestion if the human approves, or
	4. suggests one alternative
	3. narrows the selection down to a few, or
	2. The computer offers a complete set of decision/action alternatives, or
Low	1. The computer offers no assistance: human must take all decisions and actions

Note. Copyright 2000 IEEE. Repeated, with permission, from Parasuraman et al. (2000).

Tradeoffs of Locus of Authority That Changes LOAs

The term adaptive automation is commonly used to refer to technology that enables a system to flexibly recommend and/or initiate LOA changes whenever criteria for one or more triggers have been met (Oppermann, 1994; see characterizing framework in Feigh et al., 2012). Adaptive automation has been demonstrated to improve task performance and reduce attention management demands over fixed automation (e.g., Calhoun et al., 2011; Parasuraman et al., 2009). However, human acceptance of automation may decrease when the automation changes LOAs (Parasuraman & Riley, 1997). Also, the human’s situation awareness can be hampered if the adaptive changes are inappropriate, untimely, or unpredictable—often the case when complex, ill-defined situations complicate the processing involved in adaptive automation. In these cases, LOA shifts can be an annoyance and even impact system performance and safety (Miller & Hannen, 1999). Alerting the human that an LOA change is recommended or will be made by automation (mindful of human–automation exchange etiquette; Hayes & Miller, 2010) may help keep the human operator aware of the automation’s functioning, but would still impose attentional demands.

In contrast to the system recommending and/or initiating LOA changes in adaptive automation, with adaptable automation, the human retains the authority to invoke or change LOAs (Oppermann, 1994). By keeping the human in the loop, this approach “takes advantage of the human’s ability to anticipate and prepare for changes and provides the flexibility that is needed in novel situations” (van Dongen & van Maanen, 2005, pp. 2–3). Moreover, the human is able to benefit from flexible LOA assignment with less identified pitfalls associated with adaptive automation (Miller & Parasuraman, 2003, 2007). For example, adaptable automation has been found to result in higher human operator acceptance, appropriate levels of trust in the automation, and enhanced situation awareness (Parasuraman & Wickens, 2008). Adaptable automation may also facilitate recovery from automation failures, enabling the human to quickly lower LOA for direct control. In this manner, adaptable automation can help serve as a safeguard when automation is operating outside its programmed capabilities. However, having the human in charge of LOA management raises other concerns: the human could become too preoccupied to make desirable LOA changes or inadvertently select an inappropriate LOA (Kaber & Riley, 1999). Moreover, there is additional management workload associated with adaptable automation (Kirlik, 1993).

Sauer and Chavaillaz (2018, p. 117) provide a “coarse assessment” of both fixed and adaptable automation, along with three types of adaptive automation (each varying triggers). They concluded that adaptable automation is more advantageous over alternative approaches in terms of responsiveness to changing situations, automation state awareness, and perceived control over the system. These more recent conclusions match those Miller et al. (2005) made over a decade earlier based on examining tradeoffs in terms of a three-way relationship between the workload the human experiences in interacting with the system when performing tasks, the overall competency of the human–automation team across a range of circumstances, and the degree to which the automation is predictable to the human operator. In their final paragraph (Miller et al., 2005, p. 27), the authors express their skepticism on whether adaptive automation has “utility in high complexity and high criticality domains.”

Despite the documented tradeoffs, the number of publications addressing adaptive automation has consistently been about ten times that of those focused on adaptable automation. This discrepancy in publication records is also indicative of the vast investments that have been devoted to developing and researching adaptive automation in contrast to adaptable automation. In fact, very little attention has addressed adaptable automation and experiments directly comparing adaptable and adaptive automation are particularly scant. To the author’s knowledge, only seven experiments have directly compared these two flexible automation approaches. The automation in one, though, provided prompts when an LOA change should be considered (Kaber & Riley, 1999), in contrast to others. Two experiments imposed a limit on how long the automation’s assistance was available (Bailey et al., 2006; Chen et al., 2017). For the other four experiments, the flexible automation was similar: the unprompted LOA changes in the adaptable condition remained in effect until the human made a subsequent change. Table 2 summarizes the two experiments that compared adaptable automation with performance-based adaptive automation (Kidwell et al., 2012; Sauer et al., 2012; the latter also examined an event-based adaptive trigger). Results from both experiments showed adaptable (human controlled) was more favorable for LOA/task assignment than adaptive (system controlled). The other experiments (Table 3; Li et al., 2013; Ruff et al., 2018) compared performance-based adaptive automation with adaptable automation, in addition to evaluating a hybrid approach in which the adaptive automation was in effect, but the human could override and change LOA, as is possible with adaptable automation. Generally, the results were also more favorable for the adaptable automation in comparison to the adaptive automation. Additionally, many measures showed that the hybrid condition was better than adaptive automation, suggesting that further evaluation of hybrid approaches is warranted.

OPEN IN VIEWER

Table 2

Summary: Experiments Comparing Adaptive and Adaptable Automation

Citation	Kidwell et al. (2012)	Sauer et al. (2012)
Domain	multi-unmanned air vehicle simulation	process control environment: Automated Cabin Air Management System
Tasks	detect threats, allocate targets to vehicles, route vehicles, respond to chat/system status queries, analyze images	Primary: monitor 5 automatic controllers, identify/repair faults via station inputs Secondary: acknowledge messages and record pump level
Levels of Automation (LOAs)	Task: analyze images – 3 LOAs Low: no automation support Medium: automation highlights recommended option High: unless vetoed by participant within 20 s, highlighted option starts automatically	Task: identify fault and fix – 5 LOAs 1: manual, no automation support 2: automation notifies fault present 3: automation identifies fault 4: automation provides course of action 5: automation fixes fault with participant’s consent
Adaptive (System Controls LOA)	Performance-based: for 6 tasks, running count kept on frequency time to complete task violates thresholds When count exceeds ±4, LOA changes one level and count resets to 0	Performance-based: LOA determined by deviation from thresholds derived from previous experiment’s perspective memory tank level recording task Event-based: LOA determined by parameter fault type and difficulty level
Adaptable (Human Controls LOA)	Participant increase or decrease across 3 LOAs via station’s slider bar	Participant increase or decrease across 5 LOAs by selecting one of the station’s five corresponding buttons
Results	Adaptable Better Than Performance-based Adaptive Threat detection better* Decision confidence rated higher** Adaptable preferred by 8 of 12 participants	Adaptable Better than Performance-based Adaptive Workload rated less* Fatigue rated less** Anxiety rated less** Confidence rated higher* Tank log timing more accurate (ns) Task reaction time faster (ns)
		Adaptable Better than Event-based Adaptive Workload rated less* Fatigue rated less** Anxiety rated less** Confidence rated higher*

Note. *p < .05; **p < .01. ns = not significant.

OPEN IN VIEWER

Table 3

Summary: Experiments Comparing Adaptive, Adaptable, and Hybrid Automation

Citation	Li et al. (2013)	Ruff et al. (2018)
Domain	desktop robotic arm virtual environment: Basic Operational Robotic Instructional System (BORIS)	multi-unmanned air vehicle simulation
Tasks	move arm to target safely, accurately, and efficiently, detect/identify LOA changes	detect threats, allocate targets to vehicles, route vehicles, respond to system status queries, analyze images, analyze weapons
Levels of Automation (LOAs)	Task: Hazard Avoidance – 3 LOAs Low: manual, no automation support Medium: automation starts red highlight on arm when near safety limit High: automation starts red highlight & stops arm movement when near limit Task: Trajectory Control – 3 LOAs Low: manual, no automation support Medium: visualization of automation’s suggested path and hand control inputs High: same as Medium and automatically flies arm along path	Tasks: analyze images (I) and weapons (W) – 3 LOAs Low: no automation support 8 (I) or 2 (W) options Medium: automation highlights recommended option High: unless vetoed by participant within 30 s (I) or 40 s (W), highlighted option starts automatically
Adaptive (System Controls LOA)	Performance-based: LOA determined by proximity to hazards and deviation from ideal path when moving arm	Performance-based: for 4 tasks, running count kept on frequency time to complete task violates thresholds When count exceeds ±3, LOA changes one level and count resets to 0
Adaptable (Human Controls LOA)	Participant increase or decrease across 3 LOAs by selecting Manual, Medium, or High via station buttons	Participant increase or decrease across 3 LOAs via control station “>>>”and“<<<” buttons (1 LOA step per selection)
Results	Adaptable Better Than Performance-based Adaptive Mental workload rated less* Arm flight time less (ns) Adaptable preferred by 7 of 12 participants	Adaptable Better Than Performance-based Adaptive Image analysis more accurate (ns) Workload rated less (ns) Rated better task aid: 10 of 12 participants Adaptable preferred by 8 of 12 participants
	Hybrid Better Than Performance-based Adaptive Trajectory deviations smaller** Mental & physical workload rated less* Arm flight time less (ns) Hybrid preferred by 5 of 12 participants	Hybrid Better Than Performance-based Adaptive Image analysis more accurate (ns) Workload rated less (ns) Hybrid preferred by 3 of 12 participants

Note. *p < .05; **p < .01. ns = not significant.

Recommended Research

Sheridan’s statement that motivation for human robot interaction research is more “focused on building and demonstrating … than in providing detailed and validated scientific conclusions” (Sheridan, 2016, p. 531) rings true for flexible automation. To date, significant investments have been devoted to exploring adaptive automation in which the system controls the level of automation based on one or more triggers. In contrast, research addressing adaptable automation where the human controls automation level is scant. However, results from the few experiments directly comparing two flexible automation approaches unanimously favor adaptable automation (Tables 2 and 3). Moreover, given the “costs” of adaptive automation (technical sophistication/complexity involved in implementing triggers for automated LOA changes, as well as the potential for decreased human situation awareness and performance, etc.; Sauer et al., 2012), it could be viewed that the human factors community has been negligent in its focus on adaptive over adaptable automation.

Therefore, it is timely to change emphasis to experiments examining the potential contributions of adaptable automation enabled with efficient interface designs. Unless adaptable automation becomes the forefront of human–automation teaming research, the potential benefits of harnessing the unique capabilities of human and automation team members to make collaborative informed, robust, and timely decisions/system inputs may not be realized in future complex systems. Adaptable automation research is critical to evaluate its contribution toward improving system performance by augmenting the human’s pattern recognition ability and intuition/flexibility with automation’s computational/reasoning capabilities without degrading the human’s situation awareness. Specifically, systematic research that compares adaptable automation with adaptive automation for a variety of tasks/application domains, as well as examining hybrid approaches combining the two, is long overdue to address this critical gap. To help guide future flexible automation designs, research is needed to:

Results from all of these research areas may actually be applicable to the same multi-task system such that different types of flexible automation are matched to certain subsets of tasks. A spectrum of approaches might even be more appropriate given complex systems will likely be comprised of multiple types of automation, each with unique capabilities and characteristics.

The human’s workload, however, in managing a plethora of automation options would be burdensome unless mitigated, for example, by facilitating the establishment of human–automation working agreements (e.g., Gutzwiller et al., 2017). For effective working agreements to alleviate workload, research is needed to determine optimal controls and displays that enable the human–automation team to:

Additional Flexible Automation Considerations

Human–automation teaming is ripe for human factors research. Effective design of supportive controls and displays for application environments is especially critical, not only for task delegation/workload sharing, but also to exploit automation’s assistance in collaborating on courses of actions, monitoring task execution/system operation, and managing contingencies. Fortunately, progress has been made on minimizing adaptable automation’s LOA/task assignment workload by extending Playbook delegation pioneered by Miller and Parasuraman (2007) for both single human control of multiple unmanned vehicles (Calhoun et al., 2018) and single pilot-manned aircraft operation (Tokadli et al., 2021). Both enable a human operator to seamlessly transition across low-to-high LOAs and tailor automation depending on the task, situation, human’s trust in the automation, or a unique combination of these factors. These approaches are candidates for expansion to enable combined adaptable/adaptive automation support in complex systems (proposed by Miller et al., 2005), as well as determine their utility in establishing and modifying working agreements across human and automation teammates.

Research targeting flexible automation for human–automation teaming should evaluate candidate interfaces in complex, multi-task test environments, representative of actual applications including simulated or operational automation capabilities. Involvement of current operators is invaluable, both to identify requirements for the targeted domain and provide feedback on prototypes. Future research can also be informed by advancements in tools such as the human–autonomy system oversight model (Endsley, 2017), principles of effective teamwork (Joe et al., 2014), experimental design recommendations (Chen et al., 2017), and proposed measures to record (Pritchett et al., 2014; Sauer & Chavaillaz, 2018).

Key Points