Human Supervisory Control of Automation

developing Updated: 2026-06-02
supervisory-control automation monitoring human-automation-interaction teleoperation trust levels-of-automation function-allocation

Human Supervisory Control of Automation

Status: developing
Last updated: 2026-06-02
Sources: 9781119636113.Ch28.Pdf, Parasuraman 2000.Pdf
Tags: [supervisory-control, automation, monitoring, human-automation-interaction, teleoperation, trust, levels-of-automation, function-allocation]

Summary

Human supervisory control is the mode of interaction in which a person directs and oversees intelligent automated subsystems much as a supervisor directs human staff, giving directives that the automation translates into detailed action and receiving summarised feedback in return (Sheridan, 2021). The supervisor's roles fall into five time-sequential steps: planning, teaching, monitoring, intervening, and learning. Parasuraman, Sheridan & Wickens (2000) formalise the design question this raises — which functions to automate and how far — with a model in which automation applies to four classes of function (information acquisition, information analysis, decision and action selection, action implementation), each across a continuum of levels from fully manual to fully automatic. The degree of delegation is tied to how far the automation can be trusted, and high levels of automation carry human-performance costs: reduced situation awareness, complacency, and skill degradation.

Body

Context

Sheridan (2021), in his handbook chapter on human supervisory control of automation, examines the mode of interaction in which a person directs and oversees intelligent automated subsystems much as a supervisor directs human staff. He draws on his foundational work to treat supervisory control as broadly applicable to vehicles, plants, and other automated processes, and ties the degree of delegation to how far the automation can be trusted. Parasuraman, Sheridan & Wickens (2000) extend this into a design framework, addressing which system functions should be automated and to what extent. Within this knowledge base the article is the foundational entry for the automation strand: it predates and frames the broader treatments in Automation Autonomy And Ai and Human Robot Interaction, supplies the monitoring and out-of-the-loop concerns that connect directly to Situation Awareness, Mental Workload, and Vigilance And Sustained Attention, links failure detection to Human Error And Reliability, and grounds the calibration problem developed in Trust In Automation. Its concern with skill loss and complacency connects to the Ironies Of Automation critique in the artificial-intelligence-kb.

Key Points

Supervisory control derives from an analogy with human management. The term comes from the parallel between a supervisor's interaction with subordinate staff and a person's interaction with intelligent automated subsystems: a supervisor gives directives that staff translate into detailed action, while staff aggregate and transform process information into summary form for the supervisor. The degree of intelligence of the subordinate determines the supervisor's willingness to delegate and the span of the directives (Ferrell & Sheridan, 1967, cited in Sheridan, 2021). Automation permits the same interaction between a human and a process, and the concept applies broadly, including to vehicle control (PDF p. 1, orig. p. 736).

The supervisor's roles form a five-step structure: planning off-line what task to do and how to do it; teaching or programming the computer what was planned; monitoring the automatic action online to ensure all goes as planned and to detect failures; intervening to take over control or to interrupt and reprogram in emergencies; and learning from experience to do better in the future. These are usually time-sequential steps, with planning and learning performed off-line relative to the online monitoring loop (PDF p. 5, orig. p. 740).

Monitoring and trust are recurring concerns. Sheridan treats monitoring of displays and detection of failures as a major topic, including monitoring issues and the importance of prediction. Willingness to delegate depends on how much the computer can be trusted to cope with, linking trust directly to the allocation of authority between human and machine and making trust calibration a design problem rather than merely an operator attitude (PDF p. 3, orig. p. 738).

Sheridan situates supervisory control within the history of automation. From the late 1950s, computers began to intervene in control loops, relieving operators of continuous manual control such as keeping a tank at a set level while the human moved to a monitoring role (PDF p. 2, orig. p. 737). The teaching role appears in modern aircraft that pilots program rather than fly continuously (PDF p. 12, orig. p. 747).

A model for types and levels of automation (Parasuraman et al., 2000). Parasuraman, Sheridan & Wickens define automation as "a device or system that accomplishes (partially or fully) a function that was previously, or conceivably could be, carried out (partially or fully) by a human operator" (PDF p. 2, orig. p. 287). Automation is not all-or-none but varies across a continuum of levels. They adopt a 10-point scale of levels of automation of decision and action selection, in which higher levels represent increasing computer autonomy over human action — for example, the computer offering several options (low), suggesting one alternative the human still executes (intermediate), or executing automatically unless the human vetoes within a limited time (high) (PDF p. 2, orig. p. 287).

Four classes of function. The model applies these levels across a simple four-stage view of human information processing, giving four classes of function that can each be automated to differing degrees: information acquisition, information analysis, decision and action selection, and action implementation — shorthanded as acquisition, analysis, decision, and action automation, with acquisition and analysis jointly termed information automation (PDF pp. 2–3, orig. pp. 287–288). A particular system can combine automation of all four classes at different levels. Acquisition automation covers sensing, organising, highlighting, and filtering of input data; analysis automation covers prediction and integration (for example predictor and trend displays); decision automation selects among alternatives (for example the ground proximity warning system at level 4, where a single manoeuvre is recommended but can be ignored); and action automation executes the chosen response (PDF pp. 3–4, orig. pp. 288–289). Levels need not be fixed at design time — adaptive automation lets the level vary with situational demands during operation (PDF p. 4, orig. p. 289).

Evaluative criteria and human-performance costs. The appropriate level of automation should be chosen by evaluating its consequences, not by technological capability alone. Parasuraman et al. set human-performance consequences as the primary evaluative criteria and automation reliability and the costs of decision/action consequences as secondary criteria (PDF p. 4, orig. p. 289). High levels of automation, especially of decision and action functions, carry recurring costs: reduced operator situation awareness (people are less aware of state changes made by another agent than of their own), complacency (operators stop monitoring highly-but-imperfectly reliable automation and miss its occasional failures), and skill degradation with disuse. Together these produce the "out-of-the-loop" unfamiliarity that makes recovery from automation failure difficult (PDF p. 6, orig. p. 291). The model is offered as a guiding framework, not a static formula prescribing a particular type or level.

Conclusion

Sheridan (2021) concludes that supervisory control is the dominant form of contemporary human-automation interaction, structured by the five roles of planning, teaching, monitoring, intervening, and learning. Parasuraman et al. (2000) turn this into a design method: automation should be specified by type (which of the four function classes) and by level (how far along the manual-to-automatic continuum), with the choice evaluated against its human-performance consequences rather than technological feasibility. Both reach the same central challenge — matching delegation to trust — and converge on the enduring risks of out-of-the-loop performance, complacency, and skill loss as automation grows more capable.

References

Ferrell, W. R., & Sheridan, T. B. (1967). Supervisory control of remote manipulation. IEEE Spectrum, 4(10), 81–88. To be validated.

Parasuraman, R., Sheridan, T.B. & Wickens, C.D. (2000) 'A Model for Types and Levels of Human Interaction with Automation', IEEE Transactions on Systems, Man, and Cybernetics — Part A: Systems and Humans, 30(3), pp. 286–297. doi: 10.1109/3468.844354. parasuraman2000model

Sheridan, T.B. (2021) 'Human Supervisory Control of Automation', in Salvendy, G. & Karwowski, W. (eds.) Handbook of Human Factors and Ergonomics. 5th edn. Hoboken, NJ: John Wiley & Sons. sheridan2021supervisorycontrol

Open Questions

  • How should trust between human supervisor and automation be calibrated so that delegation matches the automation's actual reliability?
  • How do the five supervisory roles change as automation approaches full autonomy and the intervening role becomes rarer but more critical?