Loading field map
Human-robot interaction (HRI) began with a simple practical pressure: how should a person control a machine that is physically remote, dangerous, or delicate? The earliest answer was direct manipulation—the robot as an extension of the operator's own body. Over the next seven decades, researchers expanded the robot's role from passive tool to supervised agent, then to physical partner, and finally to social teammate. This expansion was not a clean replacement of one framework by the next. Older frameworks remain active in specialized domains, and later frameworks often absorbed or reframed the concerns of their predecessors. The result is a field in which eight major frameworks coexist, each with its own assumptions about control authority, communication modality, and the purpose of interaction.
The first framework, Teleoperation (1949–present), treated the robot as a direct extension of the human operator. In a master-slave manipulator, the operator's hand movements at a master device are replicated by a slave arm at a remote site. The defining commitment of teleoperation is continuous, low-level control: every joint motion, every grasp, every approach angle is commanded in real time. This framework was essential for handling radioactive materials in the Manhattan Project and later for undersea and space operations. But teleoperation exposed a fundamental limitation: the operator must maintain constant attention and bandwidth, which leads to fatigue, errors, and poor performance in tasks with communication delays.
Supervisory Control (1967–present) emerged directly from this limitation. Instead of commanding every motion, the operator delegates goal-level tasks to the robot, which executes them autonomously while the operator monitors progress and intervenes only when necessary. Thomas Sheridan's foundational work at NASA formalized this as a hierarchy: the human sets goals, the robot plans and executes, and the human supervises. Supervisory control preserved the human's ultimate authority while offloading the moment-by-moment burden of teleoperation. It did not replace teleoperation—direct control remains essential for tasks requiring fine dexterity or unpredictable judgment—but it introduced the idea that the robot could act as a semi-autonomous agent rather than a pure tool.
The 1990s brought a cluster of frameworks that each challenged a core assumption of earlier HRI: that the human and robot should be physically separated. Human-Robot Collaboration (1996–present) grew out of industrial manufacturing, where the goal was to place robots alongside human workers without safety cages. The key innovation was the cobot (collaborative robot), a device designed to share a workspace with humans through lightweight construction, force-limiting joints, and sensor-based speed monitoring. Collaboration emphasized safe coexistence and task-level coordination: the human and robot work on the same product, often passing parts or tools, but they do not necessarily exchange forces or negotiate control in real time.
Physical Human-Robot Interaction (pHRI, 1999–present) took the next step by focusing on intentional physical contact. Where collaboration aimed to avoid harm during accidental contact, pHRI asked how a robot could deliberately exchange forces with a human—for example, in rehabilitation therapy, where a robot guides a patient's limb through a movement. The technical core of pHRI is impedance and admittance control: the robot behaves like a spring-damper system that can be pushed or pulled by the human, rather than a rigid position-tracking device. This framework absorbed the safety concerns of collaboration but redirected them toward a positive goal: using physical interaction as a communication channel.
Shared Autonomy (1999–present) emerged from a different dissatisfaction with supervisory control. In supervisory control, authority is fixed: the human delegates, the robot executes. Shared autonomy proposed that control authority should be fluid and negotiated moment by moment. In a teleoperated vehicle, for instance, the robot might take over obstacle avoidance while the human steers, or the human might override the robot's path planner when a novel situation arises. Terrence Fong's collaborative control model, developed for vehicle teleoperation, exemplified this approach: the robot could ask for help, offer suggestions, and adjust its autonomy level based on context. Shared autonomy did not reject supervisory control's hierarchy but flattened it, making the robot an active participant in deciding who does what.
These three frameworks—collaboration, pHRI, and shared autonomy—emerged within a few years of each other and addressed overlapping concerns, but they diverged in their core commitments. Collaboration prioritized workspace coexistence and task coordination; pHRI prioritized force exchange and physical communication; shared autonomy prioritized negotiated decision-making. All three remain active today, often combined in the same system: a collaborative robot arm might use impedance control (pHRI) while sharing authority with a human operator (shared autonomy) to assemble a part (collaboration).
By the early 2000s, researchers began asking whether HRI could move beyond physical tasks and control hierarchies to include social behavior. Socially Interactive Robotics (2002–present) defined a new agenda: robots that communicate through speech, gesture, facial expression, and gaze, and that are designed to be perceived as social agents. The landmark survey by Fong, Nourbakhsh, and Dautenhahn cataloged dozens of robots—from Kismet to Cog—that used social cues to engage people in interaction. Socially interactive robotics did not replace the physical frameworks; it added a new modality. A robot could now coordinate with a human not only through forces or commands but through eye contact, turn-taking, and emotional expression.
Human-Robot Teaming (2003–present) grew out of military and disaster-response research, where robots and humans must operate as peers over extended periods. Teaming differs from collaboration in several ways. Collaboration focuses on short-term task coordination in a shared workspace; teaming emphasizes long-duration interdependence, shared mental models, and neglect tolerance—the ability of a robot to continue functioning productively when the human is not attending to it. A teaming robot must anticipate the human's needs, communicate its own state, and adjust its autonomy dynamically. This framework absorbed supervisory control's goal-level delegation and shared autonomy's negotiated authority, but added a new requirement: the robot must maintain a model of the human's attention, workload, and intentions.
Socially Assistive Robotics (SAR, 2005–present) split from socially interactive robotics by focusing on a specific outcome: behavior change in the human. Where socially interactive robots aim to be engaging conversational partners, SAR robots use social cues—praise, encouragement, gaze—to motivate people in rehabilitation, education, or therapy. The robot's goal is not just to interact but to help the human achieve a measurable improvement, such as completing more repetitions of an exercise or adhering to a medication schedule. SAR narrowed the broad agenda of social interaction to a therapeutic application, but it also expanded the evaluation criteria: success is measured not by the quality of interaction but by the human's behavioral outcomes.
Today, all eight frameworks remain active, but they have settled into different application domains. Teleoperation dominates surgical robotics (da Vinci system) and hazardous-material handling. Supervisory control is standard in drone operations and planetary rovers. Human-robot collaboration is the default in manufacturing, where cobots work alongside assembly-line workers. Physical HRI is central to rehabilitation robotics and exoskeletons. Shared autonomy is the leading framework for assistive wheelchairs and semi-autonomous vehicles. Socially interactive robotics guides the design of social robots in education and public spaces. Human-robot teaming is the framework of choice for military and search-and-rescue systems. Socially assistive robotics drives research in autism therapy, stroke rehabilitation, and elder care.
Despite this division of labor, the frameworks overlap and sometimes conflict. The most persistent debate concerns the optimal level of autonomy. Shared autonomy and teaming argue for fluid, context-dependent autonomy; supervisory control insists that the human should remain the final authority; teleoperation assumes that full human control is safest for high-stakes tasks. A second debate concerns the purpose of social behavior: socially interactive robotics treats social cues as a communication channel, while socially assistive robotics treats them as a motivational tool. A third debate revolves around evaluation: should HRI systems be judged by task efficiency, user satisfaction, safety, or long-term behavioral outcomes? Different frameworks prioritize different metrics, making cross-framework comparison difficult.
What the leading frameworks agree on is that the human should never be a passive observer. Even in teleoperation, the operator's skill matters. Even in fully autonomous systems, the human must be able to intervene. The disagreement is about how much initiative the robot should take and how that initiative should be negotiated. Shared autonomy and teaming push the robot toward proactive, context-aware behavior; supervisory control and teleoperation keep the human firmly in the loop. This tension is not a sign of immaturity—it is a productive disagreement that drives the field forward, as each framework refines its assumptions in response to the others.