Loading field map
A robot arm that welds a car body in a factory moves with precision through empty air. The same arm, asked to insert a peg into a hole, must manage forces that can jam the mechanism, break the part, or damage the arm itself. This contrast—between free-space motion and contact—defines the central challenge of robot manipulation. The history of the subfield is a sequence of frameworks that each made a different bet about how much the robot should know about its environment in advance, what it should measure during the task, and how it should decide what to do when forces at the point of contact dominate the physics.
The earliest manipulation systems sidestepped the contact problem entirely by keeping a human in the loop. Teleoperation and Supervisory Control, emerging in the 1950s for handling radioactive materials, used a master-slave architecture: a human operator moved a master arm, and a slave arm replicated the motion at a distance. The operator felt forces through mechanical linkages or, later, through force-feedback systems, making the human brain the controller. This approach worked well for hazardous environments and, later, for bomb disposal and robot-assisted surgery, where unpredictable geometry and delicate tissue demand human judgment. But teleoperation is slow, fatiguing, and limited by communication delays. The framework remains active today in safety-critical niches, but its core assumption—that a person must guide every motion—became a bottleneck as industry demanded speed and repeatability.
Model-Based Manipulator Robotics, which took shape in the 1970s, replaced the human controller with a mathematical model of the arm. Using the Lagrangian dynamics of a serial-link manipulator, engineers derived equations of motion and implemented computed-torque control: the controller calculates the torques needed to follow a desired trajectory, assuming the arm's mass, inertia, and friction are known. In free space—welding, painting, pick-and-place—this framework delivered sub-millimeter precision and cycle times no human could match. Its weakness became apparent the moment the arm made sustained contact with a rigid surface. The model assumes the environment is empty; when the arm pushes against a workpiece, unmodeled contact forces overwhelm the controller, causing instability, chatter, or breakage. Model-based control treats manipulation as a motion problem, not an interaction problem, and that assumption fails at the point of contact.
By the early 1980s, researchers recognized that manipulation tasks such as assembly, grinding, and deburring require reasoning about forces, not just positions. Two frameworks emerged in parallel, each offering a different mathematical answer to the same question: how should a robot control its interaction with a surface?
Hybrid Position-Force Control decomposes the task space into orthogonal subspaces: directions in which the robot must control position (e.g., sliding along a surface) and directions in which it must control force (e.g., pressing into the surface). The controller uses a selection matrix to switch between position and force loops, relying on a known geometric model of the contact. For a peg-in-hole insertion with a chamfered hole, hybrid control works elegantly: the robot pushes down with a commanded force while adjusting lateral position to follow the chamfer. The framework's strength is its directness—it explicitly tracks force targets—but its brittleness is severe. If the contact geometry is uncertain or changes during the task, the selection matrix becomes wrong, and the controller can drive the arm into a wall or lose contact entirely. Hybrid control assumes the environment is known well enough to precompute the subspaces.
Impedance and Admittance Control, developed in the same period, takes a fundamentally different approach. Instead of commanding force directly, the controller programs the arm to behave like a mechanical impedance—a spring-damper system with adjustable stiffness and damping. The robot measures its position and velocity, then computes the force to apply as if a virtual spring connected the actual position to a desired trajectory. In admittance control (the dual form), the robot measures force and adjusts position. The key insight is that the robot does not need to know the contact geometry; it only needs to present a compliant behavior to the environment. If the arm hits a stiff surface, it yields rather than fights. Impedance control is robust to uncertainty—it works for wiping a table, turning a crank, or opening a door without a model of the door's hinge—but it sacrifices precision in force tracking. The robot cannot guarantee a specific contact force; it can only guarantee a relationship between force and displacement. The two frameworks, hybrid and impedance, remain in productive tension. Hybrid control dominates tasks with well-known geometry and tight force tolerances (precision assembly), while impedance control dominates tasks with uncertain or varying contact (human-robot collaboration, polishing, rehabilitation robotics).
Learning-Based Manipulation, emerging in the 1990s and accelerating with deep learning after 2010, offers a third path: instead of programming a model or a control law, the robot learns a policy from data. Early work used reinforcement learning for simple tasks like juggling or grasping; later, imitation learning and deep neural networks enabled robots to learn dexterous manipulation from human demonstrations or trial-and-error in simulation. The framework's distinctive commitment is that the robot should discover its own representation of the task, often bypassing explicit dynamics models or contact models. In practice, learning-based methods rarely replace the earlier frameworks outright. They more often layer on top: a learned policy might output desired impedance parameters for a low-level impedance controller, or a learned grasp planner might feed target poses to a model-based trajectory tracker. The relationship is one of enrichment and partial absorption. Learning excels at tasks where the physics is too complex to model analytically—in-hand reorientation, cloth folding, bin picking with diverse objects—but it struggles with data efficiency, generalization to new objects, and safety guarantees. The framework is still maturing, and its current role is to handle the high-variability, low-precision end of the manipulation spectrum.
All five frameworks remain active today, and no single approach dominates the entire field. Teleoperation persists in surgery, space robotics, and bomb disposal, where human judgment is irreplaceable. Model-based control remains the workhorse of industrial robot arms for free-space motion, because it is fast, precise, and provably stable. Hybrid position-force control is the standard for precision assembly in electronics manufacturing, where the geometry is known and force tolerances are tight. Impedance and admittance control dominate collaborative robots, exoskeletons, and any application where the robot must interact safely with humans or uncertain environments. Learning-based manipulation is the fastest-growing area, driving progress in household robotics, warehouse picking, and dexterous manipulation.
What the leading frameworks agree on is that contact forces cannot be ignored: manipulation is fundamentally an interaction problem, not a motion problem. They disagree on how much prior knowledge the robot should assume. Model-based and hybrid control assume a detailed world model; impedance control assumes only a desired behavior; learning-based methods assume enough data to approximate the interaction. The deepest disagreement is about whether the robot should reason explicitly about forces (hybrid control) or treat compliance as a emergent property of a control law (impedance) or a learned policy (learning). These are not competing for a single answer; they are different tools for different tasks, and the most capable manipulation systems today combine them—using model-based planning for gross motion, impedance control for contact transitions, and learning for adaptation to novel objects. The history of robot manipulation is not a story of one framework defeating another, but of a field gradually learning that contact is not a nuisance to be eliminated but the very phenomenon that makes manipulation interesting.