Loading field map
Robot control is the engineering discipline concerned with making robots move and act in the world as intended. It sits at the intersection of mechanics, electronics, computer science, and control theory, but its defining problem is distinct: a robot is a physical system with sensors, actuators, and a computational brain, and the controller must decide what commands to send to those actuators based on sensor readings and a desired task. The central questions are deceptively simple: How does a robot know where it is? How does it decide what to do next? And how does it translate that decision into forces and torques that produce the desired motion without damaging itself or its surroundings?
The stakes are high because robots operate in the real world, where models are imperfect, sensors are noisy, and unexpected events occur. A controller that works flawlessly in simulation may fail catastrophically on hardware. The field therefore concerns itself not just with algorithms, but with robustness, stability, and the practical gap between theory and implementation.
Robot control emerged from two distinct traditions that merged in the mid-twentieth century. The first was teleoperation, developed during the nuclear and early space programs, where human operators controlled remote manipulators through mechanical or electrical linkages. The second was numerical control, where machines followed pre-programmed sequences of motions, as in early industrial milling machines. Neither tradition called itself robot control, but both contributed essential elements: the former emphasized sensing and feedback, the latter emphasized precise, repeatable trajectory execution.
The term "robot" itself, popularized by Karel Čapek's 1920 play R.U.R., originally referred to artificial laborers, not to the programmable machines that later bore the name. The connection between these early traditions and modern robot control is real but indirect. Teleoperation introduced the idea of a machine that mirrors human intent, while numerical control introduced the idea of a machine that executes a stored plan. Robot control as a distinct field crystallized when researchers began asking how to combine these ideas—how a machine could plan and execute its own motions without continuous human guidance.
The first mature approach to robot control, dominant from the 1960s through the 1980s, treated the robot as a mechanical system whose behavior could be precisely modeled. This paradigm has two layers: kinematics and dynamics.
Kinematics describes the geometry of motion without regard to forces. A robot arm's kinematic model specifies how joint angles determine the position and orientation of the end effector—the hand or tool at the tip. The forward kinematics problem is straightforward: given joint angles, compute the end-effector pose. The inverse kinematics problem is harder: given a desired end-effector pose, compute the joint angles that achieve it. Inverse kinematics may have multiple solutions, no solutions, or singularities where the robot loses degrees of freedom. Early industrial robots were often designed with kinematic structures that simplified this problem, such as intersecting joint axes that allowed closed-form solutions.
Dynamics adds forces and torques. A dynamic model of a robot arm is a set of second-order differential equations relating joint torques to joint accelerations, velocities, and positions. These equations include inertial terms, Coriolis and centrifugal forces, gravity, and friction. The central dynamic control problem is computed torque control: if the model is accurate, one can solve for the torques that produce a desired acceleration, effectively linearizing the system and allowing standard linear control techniques to be applied.
The kinematic and dynamic paradigm achieved remarkable success in industrial settings. It enabled precise, repeatable motion for welding, painting, and assembly in structured factory environments. Its limitations, however, became increasingly apparent. The models are never perfect: friction is notoriously difficult to model, payloads change, and wear alters the robot's parameters. Computed torque control degrades gracefully with model error but can become unstable if the error is large. Moreover, the paradigm assumes that the robot operates in a known, static environment, which is rarely true outside the factory floor.
Beginning in the 1980s, a different approach gained prominence, driven by the recognition that robots must operate in uncertain, dynamic environments. This paradigm emphasizes sensing and feedback over precise modeling. Instead of computing an exact trajectory offline and executing it open-loop, the robot continuously senses its environment and adjusts its behavior in real time.
The most influential formulation of this idea was the subsumption architecture, proposed by Rodney Brooks in the mid-1980s. Brooks argued that traditional AI's sense-plan-act cycle—sense the world, build a model, plan a course of action, then execute—was too slow and brittle for real-world interaction. He proposed instead a layered set of simple behaviors, each of which directly maps sensor inputs to actuator commands. Lower layers handle basic survival behaviors like obstacle avoidance; higher layers handle goal-directed behaviors like exploration. Higher layers can suppress or inhibit lower layers, but each layer operates independently and in parallel. This approach produced robots that navigated cluttered rooms and avoided obstacles without any internal model of the world.
The reactive paradigm was a genuine departure from the kinematic and dynamic tradition. It rejected the assumption that the world can be modeled accurately enough to plan in advance, and it rejected the separation between perception and action. Its strength was robustness: reactive robots degrade gracefully in the face of unexpected obstacles and sensor noise. Its weakness was the opposite of the model-based approach: reactive robots have no memory, no ability to plan ahead, and no way to reason about goals that require multiple steps. A purely reactive robot can avoid a wall but cannot navigate to a room on the other side of a building.
The field did not choose between these paradigms; it combined them. Modern robot control systems typically use a hybrid architecture: a deliberative layer that plans trajectories or tasks, a reactive layer that handles immediate feedback and safety, and an executive layer that mediates between them. This hybrid approach acknowledges that both paradigms address real problems—planning and reactivity—and that neither alone is sufficient.
A third major approach emerged in the 1990s and accelerated in the 2010s, driven by advances in optimization theory and machine learning. This paradigm treats robot control as an optimization problem: find the control sequence that minimizes a cost function subject to constraints.
Model predictive control (MPC) is the most prominent example. At each time step, MPC solves a finite-horizon optimal control problem using the current state estimate, applies the first control action, then re-solves the problem at the next time step. This receding-horizon strategy provides feedback and can handle constraints on states and inputs, such as joint limits or obstacle avoidance. MPC has become standard for legged robots, where the dynamics are complex and the robot must continuously re-plan to maintain balance.
Reinforcement learning (RL) takes a different route. Instead of using a model of the robot's dynamics, RL learns a control policy directly from experience. The robot explores its environment, receives rewards or penalties, and adjusts its policy to maximize cumulative reward. Modern deep RL methods use neural networks to represent policies, allowing them to handle high-dimensional state spaces and complex tasks. RL has achieved impressive results in simulation and in controlled laboratory settings, such as learning dexterous manipulation or agile locomotion.
The optimal and learning-based paradigm differs from its predecessors in a fundamental way. The kinematic and dynamic paradigm assumes the model is known and uses it to compute controls. The reactive paradigm abandons the model and reacts to the world. The optimization paradigm uses a model when available but frames the problem as an optimization, allowing it to incorporate constraints and trade-offs explicitly. The learning paradigm goes further, attempting to learn the model or the policy from data when the model is unknown or too complex to derive.
These approaches are not mutually exclusive. Modern systems often combine them: a learned perception module feeds a model-based controller; an MPC controller generates training data for a reinforcement learning policy; a learned residual term compensates for model error in a computed torque controller. The boundaries between paradigms have become porous, and the field increasingly views them as complementary tools rather than competing schools.
The current state of robot control is characterized by several durable trends. First, the field has largely converged on a pragmatic pluralism. Researchers and practitioners select from a toolbox of methods based on the task, the available models, and the computational resources. A factory robot arm still uses computed torque control because its environment is structured and its model is accurate. A quadrotor uses MPC because its dynamics are well understood but its environment is dynamic. A manipulation robot uses learning-based methods because grasping and dexterous manipulation resist accurate modeling.
Second, the integration of perception and control has become central. Modern robots carry cameras, lidar, and other sensors, and the controller must operate on high-dimensional sensory data. This has blurred the line between robot control and computer vision. Visual servoing, where the robot uses camera feedback to guide its motion, is now a standard technique. Learned perception modules, such as object detectors and pose estimators, feed directly into control loops.
Third, safety and verification have emerged as critical concerns. As robots move from factories into homes, hospitals, and public spaces, controllers must guarantee safe behavior even in unforeseen circumstances. This has led to a renewed interest in formal methods, reachability analysis, and control barrier functions—mathematical tools that provide certificates of safety. These methods complement, rather than replace, the optimization and learning approaches, by constraining the search space to provably safe regions.
Fourth, the field has become increasingly computational. Modern controllers often run at kilohertz rates, solving optimization problems or evaluating neural networks in real time. This has driven close collaboration between control theory and computer architecture, with specialized hardware and software frameworks designed for real-time control.
The enduring questions of robot control remain what they have always been: how to make a physical system behave as intended in a world that is never fully known. The answers have evolved from precise models, to reactive behaviors, to optimization and learning, but the fundamental tension persists. Every approach trades off model accuracy against robustness, planning against reactivity, and computation against simplicity. The field's progress has come not from resolving this tension but from developing ever more sophisticated ways to manage it.