- Atlas
- Engineering
- Robot Manipulation
Robot Manipulation
Robot manipulation is the engineering of machines that grasp, move, and alter physical objects with an arm and hand.
15 frameworks, from 1948 to today · Robotics · Engineering
- Remote manipulator 1948-Present
- Kinematic and Dynamic Modeling 1955-Present
- Teach-and-Playback Programming 1961-1990
- Dynamics and Computed-Torque Control 1970-Present
- Passive Compliance and Remote Center of Compliance 1976-1995
- Force control 1981-Present
- Grasp Mechanics and Dexterous Hands 1982-Present
- Grasp Mechanics and Force Closure 1982-Present
- Impedance Control 1985-Present
- Probabilistic Manipulation under Uncertainty 1990-Present
- Sense–Plan–Act Pipeline 1990-Present
- Imitation learning 2000-Present
- Learning-based manipulation 2010-Present
- Deep Visuomotor and End-to-End Learning 2015-Present
- Reinforcement Learning for Manipulation 2015-Present
Overview
Robot manipulation is the engineering of machines that grasp, move, and alter physical objects with an arm and hand1. Its central problem is contact: a robot arm must position itself precisely enough to touch the world, then manage the forces that arise when it does. Unlike a machine tool that cuts a fixed path, a manipulator must cope with objects whose weight, shape, and friction it may only partly know. The field's history is a sequence of attempts to answer one question—how much of the world should the machine be told about in advance, and how much should it discover through sensing and control? Each era's answer shaped what robots could do, from spraying car bodies to folding laundry.
Teleoperators and the Master–Slave Arm
The field begins not with autonomy but with distance. In the 1940s, workers handling radioactive material at the Argonne National Laboratory needed to manipulate hazardous substances behind shielding23. The mechanical master–slave manipulator, developed by Ray Goertz and colleagues and refined through the 1950s, let an operator move a handle on one side of a wall while a linked arm on the other side reproduced the motion. The "master" was the human-controlled input; the "slave" was the remote output. Mechanical linkages transmitted position directly, so the operator felt the load through the handle.
The limitation was reach and flexibility: mechanical linkages were heavy and could not span long distances. In the mid-1950s Goertz and his team at Argonne built an electrically powered, servo-controlled manipulator, the E-1, in which sensors measured the master's position and motors drove the slave to match it. This was teleoperation—remote operation through a control loop—and it introduced a problem that would recur for decades: the operator could see the slave but could not feel what it touched unless force feedback was engineered back through the master. The master–slave idea also set the vocabulary of the field: a manipulator is a chain of joints and links ending in a hand or tool, and its job is to place that hand where work must be done.
The First Industrial Arms and the Geometry of Motion
In 1961, George Devol's patent and Joseph Engelberger's company Unimation produced the Unimate, installed at a General Motors plant in New Jersey to unload die-casting machines. The Unimate was a hydraulic arm that repeated a stored sequence of joint positions4. It did not sense the world; it trusted that the world would be where it was told. This was teach-and-playback: a human physically guided the arm through a motion, the controller recorded joint angles, and the arm replayed them. It worked because the factory was engineered around the robot—parts arrived in fixed positions, and the arm's job was to be exactly repeatable.
The problem this approach could not solve was reach and orientation. A serial arm's hand position depends on all its joint angles in a complicated way, and engineers needed to compute where the hand would be, or what joint angles would put it at a desired point. The mathematics of this—kinematics—was systematized in the 1950s and 1960s. Jacques Denavit and Richard Hartenberg published in 1955 a convention for assigning coordinate frames to each link of a chain, so that the transformation from one joint to the next could be written as a standard matrix5. The Denavit–Hartenberg parameters became the common language for describing arm geometry. Given the parameters, one could solve forward kinematics—joint angles to hand position—by multiplying matrices6. The inverse problem, finding joint angles for a desired hand pose, was harder and often had multiple solutions, but closed-form solutions existed for common geometries such as the six-joint spherical wrist arm.
Dynamics came next. An arm's motion is governed not only by geometry but by inertia, gravity, and the forces between links. The equations of motion for a serial chain were derived in the 1970s and 1980s, and the computed-torque control method—using the dynamic model to calculate the torques needed for a desired acceleration—became a standard way to make an arm track a trajectory accurately. The Stanford Arm, designed by Victor Scheinman in 1969, was an early electrically driven six-joint arm built for research and used to demonstrate computer-controlled manipulation, including inserting a peg into a hole under force sensing.
Compliance, Force, and the Problem of Contact
Position control assumes the arm knows where the world is. Contact breaks that assumption: a rigid arm commanded to move into a surface will either stop or push with whatever force its motors produce, and small errors in position become large errors in force. In the 1970s and 1980s, researchers confronted this directly. The key insight was that a manipulator in contact should control force or compliance rather than position alone.
Two approaches competed. Hybrid force/position control, formalized by John Craig and Raibert in the early 1980s, divided the task space into directions where position is controlled and directions where force is controlled—for example, sliding along a surface while pressing into it7. Impedance control, developed by Neville Hogan in the mid-1980s, took a different view: rather than switching between position and force, the controller makes the arm behave like a spring-damper system with a chosen stiffness, so that contact forces arise naturally from displacement. Hogan argued that this matched how biological limbs interact with the world. The remote center of compliance, a passive mechanical device developed by Drake and others at MIT in the 1970s, solved the same problem without electronics: a compliant wrist that lets a peg self-align during insertion. These ideas made assembly tasks—peg-in-hole, turning a crank, wiping a surface—tractable for robots, and they remain the basis of force-controlled manipulation today.
Grasping and the Dexterous Hand
Parallel to arm control, a separate line of work asked how a hand should hold an object. The mechanics of grasping were studied in the 1980s and 1990s: a grasp is stable if the contact forces can resist external disturbances, and the set of forces a grasp can exert is characterized by its friction cones and force closure. Matthew Mason and J. Kenneth Salisbury's work at MIT, including Salisbury's 1982 design of the Stanford/JPL hand—a three-fingered, tendon-driven hand—connected grasp theory to hardware. The BarrettHand, a later three-fingered gripper, and the Utah/MIT hand of the 1980s explored tendon-driven, multi-fingered designs.
The gap between theory and practice was large. A human hand has many degrees of freedom and rich tactile sensing; robot hands of the era had few sensors and fragile mechanisms. Simpler grippers—two-finger parallel jaws—dominated industry because they were robust and cheap. The dexterous hand remained a research object, and the field learned that grasping is not only a geometry problem but a sensing and control problem: the hand must detect slip, modulate force, and adapt to objects it did not model.
Uncertainty, Perception, and Planning
By the 1990s, the limits of pre-programmed motion were clear. A robot that must pick objects from a bin cannot know each object's exact pose; a mobile manipulator cannot assume the world is fixed. Probabilistic methods entered manipulation from the wider field of probabilistic robotics89. Particle filters and Kalman filters estimated object poses from noisy sensors; grasp planning under uncertainty computed grasps that were likely to succeed given a distribution over object positions rather than a single estimate. The insight was that a good grasp is one that works across the range of plausible poses, not just the nominal one.
This shifted the field's center of gravity toward perception and planning. Manipulation became a pipeline: sense the scene, estimate object poses, plan a grasp and a motion, execute, and check. Each stage had its own literature, and the pipeline's fragility—errors in one stage propagated—became a central complaint. The sense–plan–act architecture, inherited from broader robotics, was powerful but slow and brittle in cluttered, changing scenes.
Learning from Data and Demonstration
The most recent turn is learning. Instead of hand-coding models of objects and contacts, researchers began to let robots learn from demonstration or from their own experience10. Learning from demonstration—recording a human's motion and generalizing it—descends from teach-and-playback but adds generalization: a policy learned from a few demonstrations can adapt to new object positions. Reinforcement learning, in which a robot improves by trial and error, became practical for manipulation in the 2010s as simulation and GPU computing made large numbers of trials feasible. Deep learning brought visual perception into the loop: a neural network can map camera images directly to grasp points or motor commands, bypassing explicit object models.
The landmark demonstrations of the 2010s and 2020s—robots learning to fold cloth, assemble objects, or pick from cluttered bins—showed that data-driven methods could handle variation that classical pipelines struggled with. The open questions are the ones the field has always faced in new form: how to generalize beyond the training distribution, how to guarantee safety when the policy is a learned black box, and how to combine the reliability of model-based control with the flexibility of learning. The history of robot manipulation is a history of moving the boundary between what is told and what is discovered—and that boundary is still being redrawn.
References
- Manipulator (device), Wikipedia
- Raymond Goertz, Wikipedia
- Remote manipulator (History), Wikipedia
- Industrial robot (History), Wikipedia
- Denavit–Hartenberg parameters, Wikipedia
- Forward kinematics, Wikipedia
- Force control (History), Wikipedia
- Partially observable Markov decision process, Wikipedia
- Monte Carlo localization, Wikipedia
- Robot learning, Wikipedia
The frameworks of Robot Manipulation
Each framework is a way of working the field has used: when it held, and the question it answers.
- 1948-Present
- 1955-Present
Kinematic and Dynamic Modeling
Where is the hand given the joint angles?
- 1961-1990
Teach-and-Playback Programming
How can a machine repeat a taught motion?
- 1970-Present
Dynamics and Computed-Torque Control
What torques move an arm along a path?
- 1976-1995
Passive Compliance and Remote Center of Compliance
How can a rigid arm fit without jamming?
- 1981-Present
- 1982-Present
- 1982-Present
Grasp Mechanics and Force Closure
When is a grasp stable against disturbance?
- 1985-Present
- 1990-Present
Probabilistic Manipulation under Uncertainty
How can a robot act without knowing poses?
- 1990-Present
Sense–Plan–Act Pipeline
How should sensing, planning and acting be ordered?
- 2000-Present
- 2010-Present
Learning-based manipulation
How can a robot learn manipulation from data?
- 2015-Present
Deep Visuomotor and End-to-End Learning
Can images map straight to motor commands?
- 2015-Present
Reinforcement Learning for Manipulation
Can a robot improve by trial and error?
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.
Dynamics and Computed-Torque Control grew out of Kinematic and Dynamic Modeling
Computed-torque control grew directly out of kinematic and dynamic modeling: the same serial-chain equations were inverted to compute joint torques for a desired acceleration. The Stanford Arm demonstrated computer control built on these models.
Carried by Victor Scheinman; Stanford Arm (1969).
Deep Visuomotor and End-to-End Learning grew out of Learning-based manipulation
Deep visuomotor learning extended learning-based manipulation by mapping camera images directly to motor commands through neural networks, bypassing explicit perception and state estimation. It is the end-to-end branch of the data-driven approach.
Grasp Mechanics and Dexterous Hands influenced by Kinematic and Dynamic Modeling
Grasp mechanics applied the same kinematic and static modeling of chains to multi-fingered hands, treating finger contacts as constraints on object motion. Salisbury's Stanford/JPL hand design used kinematic analysis of tendon-driven fingers.
Carried by J. Kenneth Salisbury; Stanford/JPL Hand (1982).
Imitation learning grew out of Teach-and-Playback Programming
Learning from demonstration grew out of teach-and-playback by recording human-guided motion, but added generalization from a few demonstrations to new situations. The recording-and-replay scheme was extended into learned policies.
Force control influenced by Dynamics and Computed-Torque Control
Force control took the dynamic model and trajectory-tracking machinery of computed-torque control and applied it to contact tasks. Craig and Raibert's hybrid force/position scheme partitioned the task space using the same model-based control framework.
Carried by John Craig; Marc Raibert.
Force control rival of Passive Compliance and Remote Center of Compliance
Passive compliance and active force control were contemporaneous answers to the same contact problem: one absorbed misalignment mechanically, the other regulated force through sensing and actuation. The remote center of compliance and hybrid force/position control competed as solutions to peg insertion.
Carried by Drake; John Craig; Marc Raibert.
Impedance Control influenced by Passive Compliance and Remote Center of Compliance
Hogan's impedance control generalized the passive-compliance idea into an actively programmed stiffness, making the arm behave like a chosen spring-damper. The mechanical compliance wrist supplied the model of contact behavior that impedance control reproduced in software.
Carried by Neville Hogan.
Impedance Control rival of Force control
Impedance control was posed against hybrid force/position control, arguing that switching between force and position modes was the wrong frame and that a single stiffness relation should govern contact. Hogan defined his approach by contrast with the force-control partition.
Carried by Neville Hogan; John Craig.
Probabilistic Manipulation under Uncertainty influenced by Grasp Mechanics and Force Closure
Grasp planning under uncertainty took force-closure criteria and evaluated them over distributions of object pose rather than a single estimate. Probabilistic methods inherited the grasp-quality measures and asked which grasps survived pose uncertainty.
Imitation learning reacted against Sense–Plan–Act Pipeline
Imitation learning was advanced as an alternative to the brittle sense–plan–act pipeline, replacing explicit pose estimation and planning with a policy learned from demonstrations. Its proponents defined it against the pipeline's error propagation.
Learning-based manipulation grew out of Imitation learning
Learning-based manipulation generalized demonstration learning into broader data-driven policy learning, including learning from the robot's own experience. The demonstration paradigm supplied the core idea of acquiring manipulation skills from data rather than hand-coded models.
Reinforcement Learning for Manipulation grew out of Learning-based manipulation
Reinforcement learning for manipulation is a branch of learning-based manipulation in which policies improve by trial and error, made feasible by simulation and GPU computing. It shares the premise that skills come from data rather than models.
Force control influenced by Remote manipulator
Teleoperation raised the problem of feeling contact through a master, and force-feedback servoing carried that concern into force control. The bilateral master–slave loop is an early instance of closing a loop on measured force.
Carried by Ray Goertz.
Impedance Control influenced by Dynamics and Computed-Torque Control
Impedance control used the dynamic model of the arm to shape its closed-loop response as a mass-spring-damper, building on the computed-torque framework. Hogan's controller presupposed the model-based torque control it modified for contact.
Carried by Neville Hogan.
Reinforcement Learning for Manipulation rival of Deep Visuomotor and End-to-End Learning
End-to-end visuomotor imitation and reinforcement learning competed as routes to learned manipulation in the mid-2010s, one learning from demonstrations and the other from reward-driven trial and error. Both bypassed explicit modeling but differed in their data source.
Beyond the field
Where Robot Manipulation's ideas came from in other fields, and what they led to there: the strongest 12 of 105 links.
Kinematic and Dynamic Modeling grew out of Kinematic Geometry in Machine Design
Robot kinematic modeling is a direct descendant of the kinematic geometry developed for machine design, taking over the treatment of planar and spatial mechanisms and extending it to open serial chains.
Dynamics and Computed-Torque Control influenced by Control Theory in Applied Mathematics
Control theory supplied the feedback, stability and tracking concepts on which computed-torque control is built; the method is a model-based application of nonlinear feedback linearization from control theory.
Dynamics and Computed-Torque Control influenced by Lagrangian mechanics in Robot Kinematics Dynamics
Lagrangian mechanics is the formalism from which manipulator dynamic equations are derived; computed-torque control inverts those equations, so the method depends directly on the Lagrangian formulation.
Carried by Joseph-Louis Lagrange.
Force control led to Operational space and force control in Robotics
Operational space control generalised force control by expressing the manipulator's task in end-effector coordinates, unifying position and force regulation.
Carried by Oussama Khatib.
Grasp Mechanics and Dexterous Hands influenced by Manipulator kinematics and dynamics in Robotics
Grasp mechanics built directly on manipulator kinematics and dynamics: grasp matrices and hand Jacobians were formulated by analogy with the manipulator Jacobian, and force-closure analysis used the same kinematic and static relations developed for serial arms.
Carried by Matthew T. Mason; J. Kenneth Salisbury; Richard P. Paul.
Grasp Mechanics and Force Closure influenced by Manipulator kinematics and dynamics in Robotics
Force-closure analysis used the manipulator kinematics and dynamics framework: grasp matrices, hand Jacobians and the mapping from finger forces to object wrenches were taken over directly from arm kinematics and statics.
Carried by Matthew T. Mason; J. Kenneth Salisbury.
Impedance Control influenced by Manipulator kinematics and dynamics in Robotics
Impedance control is formulated on manipulator kinematics and dynamics: the desired end-effector impedance is mapped through the Jacobian and the manipulator's dynamic model to joint torques.
Carried by Neville Hogan; Oussama Khatib.
Impedance Control reacted against Hybrid Force/Position Control in Robot Control
Hogan presented impedance control as an alternative to hybrid force/position control, arguing that a single unified impedance regulation replaces the switching between force and position subspaces that hybrid control requires.
Carried by Neville Hogan.
Probabilistic Manipulation under Uncertainty influenced by Probabilistic robotics in Robotics
Probabilistic manipulation under uncertainty adopted the probabilistic robotics framework: Bayesian belief representation, motion and sensor models, and particle or Gaussian filters were carried over to manipulation tasks.
Carried by Sebastian Thrun; Dieter Fox; Wolfram Burgard.
Reinforcement Learning for Manipulation influenced by Deep Reinforcement Learning in Reinforcement Learning
Deep reinforcement learning, especially the deep Q-network work of 2013-2015, supplied the algorithms that reinforcement learning for manipulation adopted for continuous robot control.
Carried by DeepMind; Volodymyr Mnih.
Remote manipulator influenced by Cybernetics and Servo-Control in Action And Motor Cognition
Cybernetics and servo-control supplied the feedback and tracking concepts that remote manipulators relied on; the operator-arm loop was treated as a servomechanism problem by the engineers who built the first master-slave arms.
Carried by Norbert Wiener.
Remote manipulator influenced by Kinematic geometry of mechanisms in Robot Kinematics Dynamics
The kinematic geometry of mechanisms, especially the analysis of linkages and degrees of freedom, gave remote manipulator designers the vocabulary for describing arm geometry and joint arrangements.

Robot Manipulation is a planet in Nousfarer
Its city has a tower for each of these 15 frameworks. Land, walk in, and climb one concept at a time.
Land on Robot ManipulationRelated fields
Fields its ideas are linked to
- Robotics 27 links
- Robot Kinematics Dynamics 19 links
- Robot Control 15 links
- Robot Learning 11 links
- Mechatronics 5 links
- Robot Motion Planning 4 links
- Mechanical Dynamics 3 links
- Artificial Intelligence 2 links
- Action And Motor Cognition 2 links
- Computer Animation 2 links
- Physical Simulation 2 links
- Human Robot Interaction 2 links
More in Robotics
- Human Robot Interaction
- Mobile Robotics
- Multi Robot Systems
- Robot Control
- Robot Kinematics Dynamics
- Robot Learning
- Robot Motion Planning
- Robot Perception And Slam
- Robotics
See where Robot Manipulation sits on the Map of Human Knowledge