Robot Control

Robot control is the part of robotics that decides how a machine should move to do what it is meant to do.

9 frameworks, from 1940 to today · Robotics · Engineering

still a working view ended · each leads to its entry below

Overview

Robot control is the part of robotics that decides how a machine should move to do what it is meant to do1. Kinematics and dynamics describe what a robot's body can do; motion planning decides where it should go; control is the layer that turns those intentions into motor commands, second by second, and keeps the machine on course when the world pushes back. The field's central problem is deceptively simple to state: a robot's actual motion never matches the ideal one, because of friction, gravity, flexibility, sensor error, and contact with objects. Every major approach in the history of robot control is an answer to that gap between the intended motion and the real one.

Servomechanisms and the feedback idea

The ancestor of all robot control is the servomechanism: a system that measures its own output and corrects the error between what it should be doing and what it is doing. The idea is old—float regulators for steam engines go back to the eighteenth century—but it became an engineering discipline in the mid-twentieth century, driven by fire-control systems, aircraft autopilots, and radar tracking during and after the Second World War. The classic tool is PID control (proportional–integral–derivative): the controller pushes in proportion to the current error, in proportion to the accumulated past error, and in proportion to how fast the error is changing2. Each term corrects a different failure—steady offset, persistent drift, overshoot—and the three gains are tuned until the loop behaves.

The first industrial robot, Unimate, installed at a General Motors plant in 1961, was in essence a set of hydraulic servos playing back recorded joint positions. For decades, industrial arms were controlled joint by joint with independent feedback loops, and this remains the backbone of factory automation. The limitation is that joint-by-joint control knows nothing about the task: it holds each joint at a commanded angle without regard to what the hand is doing in space or what forces it meets.

Computed torque and inverse dynamics

As arms became faster and more precise, engineers confronted the fact that a robot arm is a coupled, nonlinear dynamical system: moving one joint loads all the others, and gravity and inertia change with configuration. The computed-torque method, developed in the 1970s and 1980s (building on the "resolved acceleration" ideas of Luh, Walker, and Paul in 1980), attacks this directly. Using the robot's dynamic equations, the controller calculates the torques needed to produce a desired acceleration—the inverse dynamics problem—and adds a feedback correction for whatever error remains34. In effect, it cancels the robot's own complicated physics so that a simple linear controller can steer it. This "feedback linearization" made high-speed, accurate trajectory tracking possible and became the textbook foundation of manipulator control.

Adaptive control: learning the machine's own parameters

Computed torque assumes the dynamic model is accurate. Real arms have unknown payloads, uncertain friction, and parameters that drift. Adaptive control, developed for robots mainly in the 1980s, estimates those parameters online and adjusts the control law as it goes. The classic result is the work of Slotine and Li (late 1980s) on adaptive robot control, which showed that a controller could track a trajectory while continuously identifying the robot's inertial parameters. The appeal is that the machine tunes itself; the difficulty is that adaptation can be fooled by disturbances and can interact badly with unmodeled effects.

Robust control and sliding modes

Where adaptive control tries to learn the uncertainty, robust control tries to tolerate it56. Sliding-mode control, imported into robotics in the 1980s, drives the system onto a chosen surface in its state space and then keeps it there with a switching law that is insensitive to a broad class of disturbances and modeling errors. Its strength is guaranteed behavior under bounded uncertainty; its cost is chattering, the high-frequency switching that can wear actuators, which motivated later refinements such as boundary layers and higher-order sliding modes. Robust and adaptive ideas are often combined, and both remain standard tools.

Operational space and task-space control

A parallel line of work asked a different question: why control joints at all, when the task is defined by the hand? In the 1980s Oussama Khatib formulated operational space control, which describes the arm's dynamics in the coordinates of the end-effector rather than the joints, so that the controller directly regulates the hand's motion and the forces it applies. This made it natural to specify tasks in the space where they actually live—move the tool here, push with this force—and it unified motion and force control in one framework. Task-space formulations became the standard way to think about redundant arms, where many joint configurations produce the same hand pose and the extra freedom can be used for obstacle avoidance or posture control.

Impedance and hybrid force/position control

The 1980s also forced the field to reckon with contact. When a robot's tool touches a surface, position control fights the environment: commanding a position inside a rigid object produces enormous forces. Two influential answers emerged. Hybrid force/position control, associated with Raibert and Craig (1981), splits the task directions—control position along some axes and force along others, as when sliding a tool across a surface while pressing into it7. Impedance control, formulated by Neville Hogan (1985), takes a different view: rather than commanding force or position, the controller makes the robot behave like a mass–spring–damper, so it yields appropriately when it meets an object. This "make the robot compliant in software" idea became central to manipulation, assembly, and later to physical human–robot interaction, and it underlies the compliant behavior of modern collaborative robots.

Behavior-based and reactive control

Mobile robotics in the 1980s raised a challenge the arm-control tradition did not address: how to act in a messy, unpredictable world with imperfect sensing. Rodney Brooks's subsumption architecture (1986) argued against the prevailing sense–plan–act pipeline, proposing instead layers of simple reactive behaviors—avoid obstacles, wander, follow—each wired directly from sensors to actuators, with higher layers subsuming lower ones89. Brooks's lab demonstrated this on small mobile robots that moved robustly through cluttered rooms without any world model. Behavior-based control made reactivity respectable and influenced robot architectures for years, though it struggled with tasks requiring memory or long-range planning, and later systems blended reactive layers with deliberative ones.

Learning enters the loop

From the 1990s onward, control and machine learning converged. Reinforcement learning was applied to robot tasks, and learning-based methods began to supplement or replace hand-derived models—learning inverse dynamics from data, learning policies for manipulation, and later using deep networks to map sensor readings directly to actions. The attraction is that a learned controller can capture effects too complex to model; the open problems are data hunger, safety guarantees, and generalization beyond the training conditions. Learning now sits alongside classical control rather than replacing it: many modern systems use learned components inside a feedback loop whose stability still rests on classical analysis.

Where the field stands

Today's robot control is a layered practice. Joint-level PID and servo loops remain everywhere; computed-torque and inverse-dynamics methods underpin high-performance arms; operational-space and impedance control govern how manipulators interact with objects and people; robust and adaptive techniques handle uncertainty; and learning-based control is increasingly used where models fail. The questions still open are the old ones in new form: how to guarantee safety when the controller learns, how to handle contact-rich tasks that defeat clean models, and how to make controllers that generalize across robots and environments. The history of the field is a record of successive attempts to close the gap between the intended motion and the real one—and that gap is still where the work is.

References

  1. Robot control, Wikipedia
  2. PID controller, Wikipedia
  3. Inverse dynamics, Wikipedia
  4. Computed torque control, Wikipedia
  5. Adaptive control, Wikipedia
  6. Robust control, Wikipedia
  7. Force control (History), Wikipedia
  8. Behavior-based robotics (History), Wikipedia
  9. Subsumption architecture, Wikipedia

The frameworks of Robot Control

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1940-Present

    PID controller

    How do we correct error with feedback?

    Fly to its tower in Nousfarer

  2. 1975-Present

    Computed torque control

    What torques produce a desired motion?

    Fly to its tower in Nousfarer

  3. 1980-Present
  4. 1981-Present

    Hybrid Force/Position Control

    How do we control force and position together?

    Fly to its tower in Nousfarer

  5. 1985-Present

    Adaptive Control

    How can a machine tune itself?

    Fly to its tower in Nousfarer

  6. 1985-Present

    Impedance Control

    How should a robot yield to contact?

    Fly to its tower in Nousfarer

  7. 1985-Present
  8. 1986-2000

    Behavior-based robotics

    How can simple reactions produce robust behavior?

    Fly to its tower in Nousfarer

  9. 1990-Present

    Learning-Based Control

    Can a machine learn to control itself?

    Fly to its tower in Nousfarer

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

PID controllerComputed torque controlRobust Control and Sliding ModesHybrid Force/Position ControlAdaptive ControlImpedance ControlOperational Space and Task-Space Cont…Behavior-based roboticsLearning-Based Control

Beyond the field

Where Robot Control's ideas came from in other fields, and what they led to there: the strongest 12 of 45 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Robot Control is a planet in Nousfarer

Its city has a tower for each of these 9 frameworks. Land, walk in, and climb one concept at a time.

Land on Robot Control

Related fields

Fields its ideas are linked to

More in Robotics

See where Robot Control sits on the Map of Human Knowledge