Robotic control is the discipline of designing algorithms and hardware that regulate the motion, force, and behaviour of robotic systems to achieve desired trajectories or task objectives in the presence of dynamic uncertainty, environmental disturbances, and physical constraints. It spans classical feedback control strategies such as PID, computed-torque, and impedance control, through to model-predictive and learning-based controllers that adapt online. Robotic control integrates kinematics, dynamics, estimation, and optimisation to translate high-level task plans into actuator commands. It is a foundational competency enabling manipulation, locomotion, and human-robot interaction across industrial and service robot applications.
Content
- The theoretical foundations of robotic control draw from classical control engineering, with PID controllers governing individual joints since the earliest industrial manipulators in the 1960s. The Raibert planar hopper (1980s) and CMU Leg Lab work demonstrated that dynamic locomotion required more sophisticated model-based approaches, prompting the development of computed-torque control, operational-space control, and the systematic application of Lagrangian dynamics to manipulator modelling.
- Modern robotic control architectures are hierarchical. A task planner generates Cartesian waypoints or wrench profiles; a trajectory generator interpolates dynamically feasible paths; an inverse kinematics or model-predictive controller converts desired states into joint commands; and low-level servo loops close around joint position, velocity, and torque measurements at kilohertz rates. Observers such as Kalman filters and disturbance estimators reconstruct unmeasured states and compensate for unmodelled dynamics.
- Human-robot collaboration has elevated the importance of force and impedance control, which allow robots to interact safely with humans and compliant environments by regulating mechanical impedance rather than rigid position trajectories. Visual servoing integrates camera feedback directly into the control loop, enabling precise manipulation of objects with uncertain pose.
- In 2024-2025, data-driven control policies trained through reinforcement learning and imitation learning are achieving dexterous manipulation performance that classical model-based controllers could not match, particularly for in-hand manipulation of deformable objects. Diffusion-based and flow-matching policy representations are emerging as high-capacity alternatives to Gaussian mixture and regression-based policies, enabling robots to learn multimodal action distributions from human demonstrations. Integration with foundation models for language-conditioned manipulation is accelerating task generalisation.