Loading field map
Robot control is built on a persistent question: how should a machine translate its goals into coordinated motion, especially when the environment is uncertain or the robot must physically interact with objects and people? Over six decades, control frameworks have shifted from simple joint-level corrections to sophisticated approaches that combine models, learning, and hierarchical task management. Each framework emerged to address what earlier approaches could not handle, and several remain in productive coexistence today.
Joint-Space Servo and PID Control formed the first generation of practical robot control. In the early 1960s, industrial manipulators like the Unimate used simple proportional-integral-derivative (PID) feedback at each joint. A PID controller compares the desired joint angle with the measured angle and applies a correction proportional to the error, its integral, and its derivative. This approach is straightforward and requires minimal computation, which made it suitable for the first factory robots. However, PID control treats each joint independently, ignoring the robot’s coupled dynamics—the fact that moving one joint affects the others. When a robot carries a heavy load or moves quickly, the uncorrected dynamic coupling causes tracking errors and oscillations.
Model-Based Inverse-Dynamics and Computed-Torque Control (1980) addressed these limitations by explicitly incorporating a dynamic model of the robot. Instead of treating each joint in isolation, computed-torque control computes the torques needed to produce a desired acceleration, canceling nonlinear gravitational and Coriolis effects. This framework transformed robot control from a purely feedback-driven method into one that uses feedforward model information. In practice, computed-torque control dramatically improves tracking accuracy for high-speed motions, but it relies on an accurate model of the robot’s mass, inertia, and friction—an assumption that becomes problematic when the robot interacts with unknown objects or environments.
These two early frameworks still coexist: PID remains the workhorse for simpler tasks because it requires no model, while computed-torque is used when precision and speed are critical. Together they define a core trade-off between simplicity and model-based accuracy that later frameworks would revisit.
The 1980s brought a new challenge: robots that must physically interact with their surroundings—grinding a surface, assembling parts, or walking. Pure motion control is insufficient when contact forces must be regulated.
Hybrid Position-Force Control (1981) was the first framework to explicitly separate a task into position-controlled and force-controlled directions. In a peg-in-hole assembly, for instance, the robot controls force along the insertion axis while controlling position in the perpendicular plane. This approach gave robots the ability to perform contact tasks that earlier methods could not manage. However, hybrid control requires the user to predefine which directions are position- and which are force- controlled, limiting flexibility.
Robust and Sliding-Mode Robot Control (1983) emerged from the realization that computed-torque control fails when model parameters are uncertain. Sliding-mode control forces the robot’s state onto a stable surface by switching control gains, making it highly robust to disturbances and modeling errors. The trade-off is chattering—rapid switching that can excite unmodeled high-frequency dynamics. Robust control remains an important tool when hard bounds on uncertainty exist, but it does not adapt its behavior over time.
Impedance and Admittance Control (1984) offered a different philosophy. Instead of switching between position and force control, it controls the robot’s dynamic behavior—its mechanical impedance—at the point of contact. By setting a desired mass-spring-damper relationship, the robot can softly interact with the environment without needing to explicitly switch between control modes. Admittance control inverts the idea: it measures force and adjusts position. This framework is widely used in human-robot interaction and haptic devices because it makes the robot naturally compliant.
Adaptive Robot Control (1987) took a different approach to model uncertainty: it updates the dynamic model online as the robot moves. An adaptive controller estimates unknown parameters (like link mass or load) and adjusts the computed torque accordingly. This allows the robot to maintain performance even when the environment changes. Adaptive control was a direct response to the fragility of non-adaptive model-based methods, but it requires persistent excitation—enough movement to identify parameters—and can become unstable if adaptation gains are poorly chosen.
Operational-Space and Task-Space Control (1987) shifted the control objective from joint-space coordinates to the coordinates of the robot’s end-effector or any point on the robot. Instead of tracking joint angles, the controller directly regulates the position and force of the tool in Cartesian space. This made it natural for tasks like drawing a straight line or applying a constant force against a surface. Operational-space control also introduced the core idea of task prioritization: a hierarchy of tasks could be executed simultaneously, with lower-priority tasks projected into the null space of higher-priority ones. This insight would later become the foundation of whole-body control.
The interaction phase produced a rich set of coexisting frameworks: hybrid control for structured contacts, impedance for compliant interaction, adaptive methods for uncertain dynamics, and operational-space for task-level reasoning. Each solved a piece of the puzzle, and they remain in use today depending on the application’s needs.
The 2000s saw two major trends: using data to circumvent imperfect models, and coordinating entire robots with many degrees of freedom.
Learning-Based Robot Control (2000) emerged because even the best model-based and robust controllers struggle with unstructured environments, complex contact dynamics, or tasks that are hard to describe analytically. Learning-based control uses data—from demonstrations, trial-and-error, or simulation—to build a control policy or improve an existing one. Early methods like locally weighted learning allowed robots to correct tracking errors on the fly. More recently, deep reinforcement learning has enabled robots to learn dexterous manipulation and locomotion directly. Learning does not replace earlier frameworks; it often works alongside them. For example, a computed-torque controller might provide a baseline, and a learned component compensates for unmodeled friction or disturbances. However, learning introduces new challenges: sample efficiency, safety during exploration, and generalization to new situations.
Whole-Body and Hierarchical Control (2004) extended operational-space control to complex robots like humanoids and quadrupeds. A humanoid must simultaneously control its center of mass, foot contact forces, hand positions, and gaze—tasks that conflict because they all compete for the same joints. Whole-body control solves this by defining a strict hierarchy of tasks: ensuring balance is top priority, then walking direction, then arm movement, and so on. Lower-priority tasks are executed only insofar as they do not interfere with higher-priority ones. This framework absorbed the earlier null-space projection idea from operational-space control and made it the central mechanism for deploying multisensory, multiconstraint behaviors. Today, whole-body control is the standard for humanoid robots, while simpler robots may not need its full complexity.
Today’s leading frameworks do not compete for a single throne; they coexist because each excels in different regimes. PID control still runs in millions of industrial robots because it is cheap and reliable for repetitive tasks. Computed-torque control is preferred when a good model exists and accuracy is paramount. Impedance control is the default for robots that work alongside humans. Robust methods guard against worst-case uncertainties, while adaptive methods handle slowly changing parameters. Learning-based control is rapidly expanding into domains where models are unknown or tasks are complex, but it still struggles with safety guarantees and sample efficiency. Whole-body control unifies task priorities for high-DOF systems.
What these frameworks agree on is that control must be both reactive and predictive: feedback is necessary to handle disturbances, but feedforward or learned predictions improve performance. They disagree on how much to trust a model versus data, how to handle uncertainty (worst-case vs. adaptive vs. probabilistic), and how to allocate formal guarantees. The most productive recent work often combines frameworks—for instance, using a robust controller as a safety layer around a learned policy, or embedding impedance control within a whole-body hierarchy. This pluralism is unlikely to resolve into a single paradigm because the variety of robots and tasks demands a variety of control philosophies.