Glossary

381 terms, each defined in one or two sentences, with links to where the idea is explained properly.

A

A* search
A graph search that expands nodes in order of cost so far plus a heuristic estimate of the cost to go. With an admissible heuristic it returns the shortest path, and it is the standard planner on grids.
Absolute accuracy
How closely a robot reaches a position specified numerically in its base frame, as opposed to a position taught by jogging. Errors in link lengths, joint offsets and flexing under load make it worse than repeatability unless the arm is calibrated.
Absolute encoder
An encoder that reports a unique position code at every angle, so the position is known at power-on. Multi-turn versions also count whole revolutions.
Accelerometer
A sensor that measures specific force along its axes, which is acceleration minus gravity. When still it reads gravity, which gives roll and pitch but not yaw.
Hub: Sensors. Related: IMU, Gyroscope
Ackermann steering
Car-like steering in which the front wheels turn by different angles so all wheels roll about one common centre. Such vehicles have a minimum turning radius and cannot turn in place.
Action
A ROS 2 interface for long-running goals that sends feedback while it runs and can be cancelled, such as navigating to a pose. It is built from topics and services underneath.
Action chunking
Having a policy predict a short sequence of future actions at once rather than one action per step. It smooths motion and reduces compounding error in imitation learning.
Action space
The set of outputs a policy can choose from, such as joint torques, joint position targets or end-effector deltas. The choice strongly affects how easily a policy learns and transfers to hardware.
Actuator
A device that produces motion or force from an energy source, such as an electric motor, a hydraulic cylinder or a pneumatic muscle. A robot joint actuator usually packages a motor, gearbox, encoder and drive.
Admissible heuristic
A cost-to-go estimate that never exceeds the true remaining cost, such as straight-line distance on a map. A* with an admissible heuristic is guaranteed to find an optimal path.
Hub: Motion planning. Related: A* search
Admittance control
The counterpart of impedance control for position-controlled robots: a measured force is turned into a motion offset through a virtual mass, spring and damper. It is the common way to add compliance to stiff industrial arms with a wrist force sensor.
AGV
Automated guided vehicle, a mobile robot that follows fixed routes marked by wires, tape, magnets or reflectors. AGVs are predictable but need infrastructure changes when routes change.
Hub: Mobile robots. Related: AMR, ISO 3691-4
AMCL — see Monte Carlo localisation
Adaptive Monte Carlo localisation, which varies the number of particles with the uncertainty of the estimate.
AMR
Autonomous mobile robot, a wheeled platform that navigates with onboard sensors and maps rather than fixed guides. AMRs replan around obstacles instead of stopping.
Analytical inverse kinematics
A closed-form solution of inverse kinematics derived for a specific arm geometry, returning every solution branch directly. Many 6-axis arms with a spherical wrist have one, while general geometries need numerical methods.
ANSI/A3 R15.08
The US standard series on safety of industrial mobile robots, in three parts: requirements for the robot, for integrating robots and fleets into a site, and for end users. It is published by the Association for Advancing Automation (A3).
AprilTag
A family of square fiducial markers and detection software from the APRIL Robotics Laboratory at the University of Michigan. A detection gives the tag's ID and its 3D position and orientation relative to the camera.
ArduPilot
An open-source autopilot, licensed under GPLv3, for multicopters, helicopters, fixed-wing aircraft, rovers, boats and submarines. It is separate from PX4 but also speaks MAVLink.
Attitude control
Controlling an aircraft's or spacecraft's orientation. On a multirotor it runs as a fast inner loop beneath the slower position and velocity loops.
Autonomous mobile robot — see AMR
The full name of AMR.
Autonomy
How much of a task a robot completes without a person deciding each step. It is a scale rather than a yes or no, running from full teleoperation through supervised operation to unattended runs.
AUV
Autonomous underwater vehicle, an untethered submarine robot that runs pre-planned missions. GNSS and radio do not work underwater, so it navigates with inertial sensors, a Doppler velocity log and acoustic positioning.
Axis-angle
A rotation written as a unit axis and an angle about it. Multiplying the axis by the angle gives the three-number rotation vector used in many optimisers and in OpenCV.

B

Backdrivability
How easily a joint can be turned by pushing on its output when the motor is unpowered. Backdrivable joints are safer in contact and allow torque sensing from motor current.
Backlash
Free play between meshing gear teeth, so the output can move a little before the input engages it. It causes position error and can make a position loop oscillate.
Behaviour cloning
The simplest form of imitation learning, which trains a policy with supervised learning to output the demonstrated action for each observation. Small errors can push the robot into states the data never covered.
BLDC motor
A brushless DC motor, which switches current between its windings electronically instead of through brushes. It is the standard choice for robot joints because it is efficient, compact and needs little maintenance.
Brushed DC motor
A motor whose current is switched mechanically by brushes rubbing on a commutator. It is cheap and easy to drive, which makes it common in hobby robots, but the brushes wear out.
Brushless DC motor — see BLDC motor
The full name of a BLDC motor.
Bundle adjustment
Jointly refining camera poses and 3D point positions to minimise the reprojection error of all observed features. It is the back end of most visual SLAM and structure-from-motion systems.
BVLOS
Beyond visual line of sight, flight where the remote pilot cannot see the aircraft directly. Most aviation regulators require specific approvals for it.

C

C-space — see Configuration space
Short for configuration space.
Hub: Kinematics.
Camera calibration
Estimating a camera's intrinsics and distortion, usually from many images of a checkerboard or marker board. Recalibrate after any change to focus or lens.
Camera extrinsics
The pose of a camera relative to another frame, such as the robot base or another sensor. Errors in extrinsics show up directly as errors in where the robot thinks objects are.
Camera intrinsics
The parameters inside the camera model: focal lengths in pixels, principal point and distortion coefficients. They map 3D points in the camera frame to pixel coordinates.
CAN bus
Controller Area Network, a two-wire serial bus from the automotive industry, standardised as ISO 11898, on which devices broadcast short prioritised messages. Classic CAN carries up to 8 data bytes per frame and CAN FD up to 64.
Capture point
The point on the ground where a legged robot would need to step to come to rest, based on its current centre of mass position and velocity. Push recovery and step planning use it.
Centre of mass
The point where a body's mass can be treated as concentrated for translational motion. Balance control for legged robots keeps the centre of mass's ground projection, or its dynamic equivalent, inside the support region.
Centre of pressure
The point where the resultant ground reaction force acts on a foot or set of feet, measured with foot force sensors. On flat ground it coincides with the zero moment point when the robot is balanced.
Centroidal dynamics
The dynamics of a robot's total linear and angular momentum about its centre of mass, driven only by external contact forces and gravity. It is a reduced model used for planning legged motion.
Cobot
A collaborative robot, an arm designed to work near people without fixed guarding, usually through power and force limiting. Whether a particular application is safe without guarding still depends on its risk assessment, not on the arm alone.
colcon
The command-line build tool for ROS 2 workspaces, which builds many packages in dependency order. colcon build followed by sourcing the install setup file is the usual workflow.
Collaborative operation
Robot operation in which people and the robot share a workspace while the robot is running, using specific safeguarding methods. The standard methods are safety-rated monitored stop, hand guiding, speed and separation monitoring, and power and force limiting.
Collision checking
Testing whether a robot configuration, or the motion between two configurations, intersects obstacles or the robot itself. It usually dominates the run time of sampling-based planners.
Commissioning
Installing, configuring and proving a robot system on site until it runs as specified. It includes tool and frame calibration, safety validation and handover to the operators.
Complementary filter
A simple fusion filter that high-passes one signal and low-passes another, for example trusting the gyroscope over short times and the accelerometer over long ones. It is a cheap way to estimate tilt on a microcontroller.
Computed-torque control
A controller that uses the robot's dynamic model to cancel its nonlinear terms, leaving a simple linear system for a PD loop to stabilise. It works well when the model is accurate and the joints accept torque commands.
Configuration space
The space of all possible joint values of a robot, with one dimension per degree of freedom. Motion planners search this space, where the robot is a single point and obstacles become forbidden regions.
Contact model
How a simulator computes forces when bodies touch, including penetration, friction and restitution. Contact modelling is the largest source of differences between simulators and between simulation and reality.
Contact schedule
The planned timing of when each foot is in contact with the ground or in the air. Model predictive controllers for legged robots often take it as a fixed input.
Contact-rich manipulation
Tasks that depend on sustained or changing contact, such as insertion, wiping or assembly. They need force control or compliance rather than pure position control.
Control bandwidth
The frequency up to which a closed loop tracks its setpoint well. Inner current loops have higher bandwidth than velocity loops, which are faster than position loops.
Control rate
How often a controller reads its sensors and updates its output, given in hertz. Torque-controlled arms commonly run joint loops at 1 kHz, while high-level planners update far less often.
Coordinate frame
An origin and three perpendicular axes attached to something, such as the world, a robot base, a camera or a tool. Positions and orientations only mean something when you say which frame they are expressed in.
Coriolis and centrifugal terms
The velocity-dependent torques in a robot's equations of motion that arise because links move in rotating frames. They grow with the square of joint speeds and matter most in fast motions.
Costmap
A grid in which each cell holds a cost for the robot to occupy it, built from the static map, live sensor data and an inflation layer around obstacles. Nav2 planners and controllers plan on costmaps.
Covariance
A matrix that describes the uncertainty of an estimate and how errors in its components are correlated. ROS pose and odometry messages carry a 6 by 6 covariance for this reason.
Cycle time
The time a robot or cell takes to complete one repetition of its task, which sets throughput. Integrators budget it move by move.
Cycloidal drive
A reducer in which a lobed disc rolls inside a ring of pins, giving high ratios and high shock-load capacity. It is common in the larger joints of industrial arms.

D

DAgger
Dataset aggregation, an imitation learning method that runs the learned policy, asks an expert to label the states it reaches and retrains on the growing dataset. It reduces errors from distribution shift.
Damped least squares
A numerical inverse kinematics update that adds a small damping term to the Jacobian pseudo-inverse. It trades a little tracking accuracy for bounded joint speeds near singularities.
Data association
Deciding which map landmark, if any, a new observation belongs to. Wrong associations are a main cause of SLAM failures.
DDS
Data Distribution Service, an Object Management Group standard for publish-subscribe middleware with discovery and quality-of-service settings. ROS 2 uses DDS implementations as its default communication layer, and the RTPS wire protocol lets different vendors' implementations talk to each other.
Dead reckoning
Working out current position from a known start, heading, speed and elapsed time, without external fixes. Odometry is dead reckoning with robot sensors.
Hub: State estimation. Related: Odometry, IMU
Degrees of freedom
The number of independent values needed to describe a robot's configuration, which for an arm is usually the number of joints. A rigid body moving freely in space has 6: three for position and three for orientation.
Demonstration
A recorded example of a task, usually synchronised observations and actions captured while a person teleoperates the robot. Data quality and consistency matter more than raw quantity.
Denavit-Hartenberg parameters
A convention that describes each link of a serial chain with four numbers: link length, link twist, link offset and joint angle. Two versions exist, standard and modified, and they place frames differently, so check which one a table uses.
Depth camera
A camera that reports distance per pixel, using stereo matching, structured light or time of flight. Most also give a colour image aligned to the depth, called RGB-D.
Depth image
An image whose pixel values are distances rather than colours. With the camera intrinsics it converts directly into a point cloud.
Dexterous hand
A multi-fingered robot hand with many actuated joints, able to reposition objects within the grasp. Control and sensing are much harder than for a two-finger gripper.
DH parameters — see Denavit-Hartenberg parameters
Short for Denavit-Hartenberg parameters.
Hub: Kinematics.
Differential drive
A drive with two independently powered wheels on one axle, steered by driving them at different speeds, plus casters for support. It can turn on the spot but cannot move sideways.
Diffusion policy
A policy that generates actions with a diffusion model, starting from noise and refining it step by step into an action sequence. It handles demonstrations in which the same situation was solved in several different ways.
Digital twin
A simulation model of a specific physical system that is kept in step with it using live data, so it can be used for monitoring, testing changes or planning. Many products sold as digital twins are plain offline simulations, so ask what data keeps the model current.
Dijkstra's algorithm
A graph search that expands nodes in order of cost from the start and finds shortest paths to all reachable nodes. It is A* with a heuristic of zero.
Distribution shift
A mismatch between the situations a policy saw in training and those it meets in use. In imitation learning the policy's own errors cause it, by leading into states the demonstrations never visited.
DoF — see Degrees of freedom
Short for degrees of freedom, as in a 6-DoF arm.
Domain randomisation
Varying simulator parameters such as masses, friction, latency, lighting and textures during training so the policy works across all of them. The real robot then looks like one more variation.
Drone — see UAV
The everyday term for an uncrewed aerial vehicle.
DWA — see Dynamic window approach
Short for dynamic window approach.
Dynamic window approach
A local planner that samples velocity commands reachable within the next control period, simulates each briefly and picks the best by a score of progress, clearance and speed.

E

E-stop — see Emergency stop
Short for emergency stop.
East-north-up (ENU)
A local world frame with x pointing east, y north and z up. ROS uses ENU for local geographic frames, while aviation and PX4 often use north-east-down.
EKF — see Extended Kalman filter
Short for extended Kalman filter.
Emergency stop
A manually operated control, usually a red mushroom button on a yellow background, that stops hazardous motion and stays latched until deliberately reset. It is a complementary measure, not a substitute for safeguarding.
Encoder
A sensor that measures the angle or position of a shaft, usually optically or magnetically. Joint encoders are what make a robot's repeatability possible.
End-effector
The device at the tip of a robot arm that interacts with the task, such as a gripper, suction cup, welding torch or screwdriver. The word is also used for the frame attached to that device.
End-of-arm tooling
The industrial term for whatever is mounted on the robot flange, such as grippers, tool changers, sensors and process tools. It is often designed per application by the integrator.
ENU — see East-north-up (ENU)
Short for east-north-up.
EOAT — see End-of-arm tooling
Short for end-of-arm tooling.
Epipolar geometry
The geometric relation between two views of the same scene, which restricts a point's match in the second image to a line. Stereo matching and two-view pose estimation use it.
Equations of motion
The differential equations that relate joint torques to joint accelerations, velocities and positions. For an arm they take the form mass matrix times acceleration plus Coriolis, centrifugal and gravity terms equals applied torque.
EtherCAT
An industrial Ethernet fieldbus, originally developed by Beckhoff Automation and covered by IEC 61158, in which one frame passes through every device and each reads and writes its data on the fly. Its short, deterministic cycle times suit multi-axis robot joints.
Euler angles
Three angles that describe an orientation as a sequence of rotations about chosen axes, for example z then y then x. There are 12 possible axis orders and two ways to apply them, so always state the convention.
Extended Kalman filter
A Kalman filter for nonlinear systems that linearises the motion and measurement models around the current estimate using their Jacobians. It is the workhorse of robot sensor fusion but can diverge if the linearisation is poor.

F

Factory acceptance test
A test of a robot system at the builder's premises, witnessed by the customer, before it ships. A site acceptance test repeats the checks after installation.
FAT — see Factory acceptance test
Short for factory acceptance test.
Feature matching
Detecting distinctive points in images, describing their surroundings and pairing points that look alike across images. It underlies visual odometry, SLAM and stereo.
Feedback control
Measuring the output of a system, comparing it with the desired value and adjusting the input to reduce the difference. Almost every robot joint runs a feedback loop on position, velocity or current.
Feedforward control
Adding a command computed from the desired motion or a known disturbance, without waiting for an error to appear. Gravity and inverse dynamics terms are common feedforward signals on robot arms.
Fiducial marker
A printed pattern designed to be detected reliably and to give its ID and pose from one camera image. Robots use them for calibration, docking and ground truth.
Field robotics
Robots that work in unstructured outdoor settings such as farms, mines, forests, construction sites and the sea. Weather, terrain and weak communications drive most of their design.
Field-oriented control
A method for driving brushless motors that transforms phase currents into a rotating frame aligned with the rotor, then controls the torque-producing current directly. It gives smooth torque and is the usual basis for torque-controlled joints.
Fixed-wing UAV
An uncrewed aircraft that gets lift from wings, giving much longer range and endurance than a multirotor. It needs space to launch and land and cannot hover.
FK — see Forward kinematics
Short for forward kinematics.
Hub: Kinematics.
Flight controller
The onboard computer and firmware that stabilise an aircraft, fuse its IMU and other sensors, and run the attitude, rate and position loops. PX4 and ArduPilot are the main open-source firmware.
Floating base
A robot whose base is not bolted to the world, such as a legged robot or humanoid, so its model has six extra unactuated degrees of freedom for the base pose. Only contact forces can move those six.
FOC — see Field-oriented control
Short for field-oriented control.
Force closure
A property of a grasp whose contacts, with friction, can resist any external force and torque on the object. It is a standard test of grasp quality.
Force control
Regulating the force a robot applies to its environment rather than its position. Hybrid schemes control force along some axes and position along the others, as when wiping a surface.
Force-torque sensor
A sensor, usually mounted at the wrist, that measures the three forces and three torques passing through it. It enables admittance control, contact detection and payload estimation.
Forward dynamics
Computing the accelerations that result from given joint torques and external forces. Simulators run forward dynamics at every time step.
Forward kinematics
Computing the pose of the end-effector from the joint positions. For a serial arm it is a chain of homogeneous transformations, one per joint, and always has a single answer.
Friction cone
The set of contact forces that do not cause slipping, a cone around the surface normal whose half-angle is set by the friction coefficient. Grasp analysis and legged robot controllers both keep contact forces inside it.
Friction model
A description of resisting torque in a joint, usually a constant Coulomb part that opposes motion plus a viscous part proportional to speed. Gearboxes make joint friction large and hard to predict, which limits torque estimation from motor current.
Functional safety
The part of overall safety that depends on control systems, including their software, responding correctly to faults and hazardous conditions. It is shown through a defined lifecycle and quantified reliability targets.

G

Gait
The pattern and timing of leg movements, such as walk, trot, pace or bound for a quadruped. Gaits differ in how many feet are on the ground at once and therefore in stability and speed.
Gazebo
An open-source robot simulator from Open Robotics with tight ROS 2 integration, sensor simulation and plugins. The older Gazebo Classic reached end of life in early 2025, and current releases are named Gazebo.
Gear ratio
How many turns of the input shaft produce one turn of the output. A high ratio multiplies torque and divides speed, and it multiplies reflected inertia by the ratio squared.
Geofence
A virtual boundary that a vehicle is not allowed to cross, enforced by its autopilot with a preset action such as hold or return. It is a basic safety layer for drone operations.
Gimbal lock
The loss of one rotational degree of freedom in an Euler angle description when the middle rotation aligns the first and third axes, for example at a pitch of 90 degrees. The orientation is still valid, but the angles stop being unique and their rates blow up.
Global planner
The part of a navigation stack that plans a full route from the robot's position to the goal on the map. It runs at a low rate and usually uses A* or a similar graph search.
GNSS
Global navigation satellite systems, the umbrella term for GPS, Galileo, GLONASS, BeiDou and regional systems. A plain receiver gives a position good to a few metres outdoors and nothing useful indoors.
GPS — see GNSS
The United States satellite navigation system, often used loosely for any GNSS receiver.
Hub: Sensors.
Grasp
A set of contacts between a gripper and an object, defined by the gripper's pose and finger positions, that holds the object. A good grasp resists the forces the task will apply.
Grasp planning
Choosing where and how to grasp an object, from a known model or directly from sensor data. Learned grasp planners score many candidate gripper poses on a point cloud or depth image.
Gravity compensation
Adding the torques needed to hold the arm against gravity to the controller output, computed from the dynamic model. With it, a torque-controlled arm floats in place and can be pushed around by hand.
Gripper
An end-effector that holds objects, by closing fingers, by suction or by magnetism. Most industrial picks use either a parallel-jaw or a vacuum gripper.
Gyroscope
A sensor that measures angular velocity about its axes. Integrating it gives orientation change, which drifts over time because of bias.

H

Hand guiding
A collaborative method in which an operator moves the robot directly by hand through a guiding device at the end-effector. It is used for teaching positions and for assisted lifting.
Hand-eye calibration
Finding the fixed transform between a camera and the robot, either a camera on the wrist or one mounted in the cell. It is solved from several robot poses that each view a calibration target.
Hardware-in-the-loop
Testing real controller hardware against a simulated robot or plant that runs in real time. It catches timing and interface faults before the controller drives real motors.
Harmonic drive (strain wave gear)
A compact gear that uses a flexible toothed cup, deformed by an elliptical wave generator, to mesh with a rigid ring, giving high ratios with essentially no backlash in one stage. Harmonic Drive is a registered trademark, and the generic name is strain wave gearing.
HIL — see Hardware-in-the-loop
Short for hardware-in-the-loop.
Holonomic robot
A mobile base that can move in any direction in the plane and rotate independently, without manoeuvring. Omnidirectional and mecanum wheels make a base holonomic.
Homogeneous transformation
A 4 by 4 matrix that packs a rotation matrix and a translation vector so that a rigid-body motion becomes a single matrix product. Chaining frames is then matrix multiplication, and the inverse undoes the motion.
Humanoid robot
A robot with a human-like body plan, usually two legs, two arms and a head. The form lets it use spaces and tools made for people, at the cost of hard balance and control problems.
Hydraulic actuator
A cylinder or rotary motor driven by pressurised oil. It offers very high force for its size and is used on heavy machinery and some legged robots, but it needs a pump, hoses and leak management.

I

ICP — see Iterative closest point
Short for iterative closest point.
IEC 61508
The basic functional safety standard for electrical, electronic and programmable electronic safety-related systems, which sector standards such as IEC 62061 build on. It defines four safety integrity levels, SIL 1 to SIL 4.
IEC 62061
The machinery-sector functional safety standard for safety-related control systems, written within the IEC 61508 framework. It rates safety functions by safety integrity level and is an alternative route to ISO 13849.
IK — see Inverse kinematics
Short for inverse kinematics.
Hub: Kinematics.
Imitation learning
Learning a policy from demonstrations of the task, usually recorded by teleoperation. It needs far fewer trials than reinforcement learning but cannot do better than its data without extra steps.
Impedance control
Controlling the relationship between motion and contact force so the robot behaves like a chosen spring and damper around a target. It suits contact tasks and is usually implemented on torque-controlled arms.
IMU
An inertial measurement unit, which combines accelerometers and gyroscopes, and sometimes a magnetometer, to measure linear acceleration and angular rate. Integrating its output alone drifts, so it is fused with other sensors.
Incremental encoder
An encoder that outputs pulses as the shaft turns, so position is found by counting from a reference. It loses its position at power-off and needs homing.
Industrial robot
A programmable, multi-axis manipulator used in manufacturing for tasks such as welding, painting, handling and assembly. Traditional industrial robots work behind guarding, separated from people.
Inertia tensor
A 3 by 3 symmetric matrix describing how a body's mass is spread around its centre of mass, which sets how hard it is to spin about each axis. URDF and MJCF files list one per link.
Innovation
The difference between an actual measurement and the measurement the filter predicted. Large innovations compared with their expected covariance flag outliers or a wrong model.
Integral windup
Build-up of a large integral term while an actuator is saturated, which then causes a big overshoot once the error changes sign. Anti-windup schemes clamp or bleed off the integrator when the output hits its limit.
Inverse dynamics
Computing the joint torques needed to produce given accelerations at the current position and velocity. Controllers use it as a feedforward term.
Inverse kinematics
Finding joint positions that place the end-effector at a desired pose. It may have several solutions, infinitely many for a redundant arm, or none if the pose is out of reach.
Isaac Sim
NVIDIA's robot simulation framework, built on Omniverse libraries and OpenUSD, used for testing robots and generating synthetic training data. It runs on NVIDIA GPUs.
ISO 10218
The two-part international standard on industrial robot safety: Part 1 covers the robot itself and Part 2 covers robot applications and robot cells. The 2025 revision brought most of the collaborative-application content of ISO/TS 15066 into the series.
ISO 13482
The ISO standard on safety requirements for personal care robots, such as mobile servant robots, physical assistant robots and person carriers. At the time of writing a revision covering service robots more broadly is in draft.
ISO 13849
The machinery safety standard for designing safety-related parts of control systems, which rates them by performance level. Part 1 gives the general design principles.
ISO 3691-4
The ISO standard on safety requirements and verification for driverless industrial trucks and their systems, which covers AGVs and many AMRs. It does not cover rail-guided or remote-controlled trucks.
ISO/TS 15066
A 2016 ISO technical specification on safety for collaborative industrial robot operation, including power and force limiting. Much of its content has been incorporated into the 2025 edition of ISO 10218.
Iterative closest point
An algorithm that aligns two point clouds by repeatedly pairing each point with its nearest neighbour and solving for the best rigid transform. It needs a decent initial guess or it converges to the wrong alignment.

J

Jacobian
The matrix that maps joint velocities to end-effector velocity at the current configuration. Its transpose maps a tool force to joint torques, and it loses rank at singularities.
Joint
A connection between two links that allows a defined relative motion, usually rotation about one axis or sliding along one axis. Each actuated joint adds one or more degrees of freedom to the mechanism.
Joint space
Describing a robot's state or motion in terms of its joint values. A joint-space move interpolates each joint independently, so the tool path between the endpoints is curved.

K

Kalman filter
A recursive estimator for linear systems with Gaussian noise that alternates a prediction from a motion model with a correction from each measurement. It tracks both the state and its covariance, and is optimal under those assumptions.
Keyframe
A selected camera frame or scan that a SLAM system keeps in its map, while frames in between are used only for tracking. Keeping only keyframes bounds memory and optimisation cost.
Kinematic chain
A sequence of links connected by joints, from a base to a tip. An open chain has a free end, as in a typical robot arm, and a closed chain loops back on itself.
Kinematic redundancy
Having more joints than the task needs, such as a 7-DoF arm placing a tool with a 6-DoF pose. The extra freedom can be used to avoid obstacles, joint limits or singularities.

L

Lagrangian mechanics
Deriving equations of motion from the difference between kinetic and potential energy instead of from force balances. It is the usual textbook route to the manipulator equation.
Landmark
A distinct, recognisable feature of the environment, such as a corner, a pole or a fiducial marker, whose position a SLAM system estimates and reuses. Landmark-based maps are sparse compared with grids.
Launch file
A file, written in Python, XML or YAML, that starts a set of ROS 2 nodes with their parameters, remappings and namespaces. It turns a multi-node system into one command.
Lens distortion
Deviation of a real lens from the pinhole model, mainly radial distortion that bends straight lines near the image edges. Calibration estimates it so images can be undistorted.
LeRobot
An open-source PyTorch library from Hugging Face for real-world robot learning, with tools for recording datasets, training policies and running them on low-cost arms. It is released under the Apache 2.0 licence.
Lidar
A sensor that measures distance by timing pulses or modulated beams of laser light, building 2D scans or 3D point clouds. 2D lidars are standard on indoor mobile robots, 3D units on outdoor vehicles.
Lifecycle node
A ROS 2 node with managed states, unconfigured, inactive, active and finalised, and explicit transitions between them. It lets a system bring up and shut down components in a controlled order.
Light curtain
A safety device that projects a row of infrared beams across an opening and triggers a stop when any beam is broken. It guards cell entrances without a physical door.
Linear inverted pendulum model
A simplified model of walking that treats the robot as a point mass at constant height on a massless leg. Its linear dynamics make walking pattern generation and balance control fast to compute.
Linear quadratic regulator — see LQR
The full name of LQR.
Hub: Control.
A rigid body in a robot's mechanism, such as an upper arm or a shin. Links are connected to each other by joints.
LIPM — see Linear inverted pendulum model
Short for linear inverted pendulum model.
Local planner
The part of a navigation stack that turns the global route into velocity commands at a high rate, while avoiding obstacles the map did not show. Nav2 calls this the controller.
Localisation
Estimating a robot's pose within a known map. Global localisation starts with no idea of the pose, while tracking refines an estimate that is already roughly right.
Loop closure
Recognising that the robot has returned to a place it mapped before and adding that constraint to the map. It corrects the drift built up around the loop, and a false loop closure can wreck the map.
LQR
The linear quadratic regulator, which computes the optimal state-feedback gain for a linear system by minimising a weighted sum of state error and control effort. Tuning means choosing those weights rather than individual gains.
Lyapunov stability
A way to prove a system returns to equilibrium by finding an energy-like function that never increases along its motion. Many nonlinear robot controllers are justified with such a function.

M

Magnetometer
A sensor that measures the local magnetic field, used as a compass for heading. Motors, steel structures and wiring distort it, so it is unreliable on many robots and indoors.
Hub: Sensors. Related: IMU, Gyroscope
Manipulability
A scalar measure of how far a configuration is from a singularity, often the square root of the determinant of the Jacobian times its transpose. Planners use it to keep an arm in well-conditioned poses.
Mass matrix
The symmetric, positive-definite matrix in the equations of motion that relates joint accelerations to the torques needed to produce them. It changes with configuration, because the arm's inertia about each joint depends on its pose.
A lightweight messaging protocol for talking to drones and between their onboard components, with a header-only C reference library. Ground stations and companion computers use it to command PX4 and ArduPilot vehicles.
Measurement noise
The uncertainty a filter assigns to each sensor reading. The ratio between process and measurement noise decides how much the filter trusts the model versus the sensors.
Mecanum wheel
A wheel with angled rollers around its rim, so that four of them driven at chosen speeds move the base in any direction. Mecanum bases are omnidirectional but sensitive to floor quality.
Mechatronics
Engineering that combines mechanical design, electronics, control and embedded software in one product. A robot joint module with its motor, gearbox, encoder and drive board is a typical mechatronic unit.
Message
A typed data structure sent over a ROS topic, defined in a .msg file, such as sensor_msgs/Image or geometry_msgs/Twist. Services and actions have their own interface files built from the same field types.
micro-ROS
A port of ROS 2 to microcontrollers with limited resources, which connects them to a ROS 2 system on a larger computer. It lets firmware on a motor board or sensor appear as ordinary ROS 2 nodes.
Minimum-jerk trajectory
A trajectory that minimises the integral of squared jerk, which for a point-to-point move gives a fifth-order polynomial. It produces smooth starts and stops that are easy on gearboxes.
MJCF
The XML model format of MuJoCo, describing bodies, joints, geometry, actuators, sensors and simulation options in one file. MuJoCo can also load URDF and convert it.
Model predictive control
A controller that solves a short-horizon optimisation over a model at every time step, applies the first command and repeats. It handles constraints such as torque and joint limits directly, at the cost of more computation.
Monte Carlo localisation
Localisation in a known map with a particle filter. The ROS 2 navigation stack ships an adaptive version, AMCL, for 2D lidar.
Motion planning
Finding a sequence of robot configurations from a start to a goal that avoids collisions and respects the robot's limits. Planners work in configuration space.
Motor driver
The power electronics that turn a low-power command into the currents a motor needs, often with a current loop on board. Higher-end drives also run velocity and position loops and talk over CAN or EtherCAT.
MoveIt 2
The main open-source motion planning framework for ROS 2 arms, combining inverse kinematics, collision checking, planners such as those in OMPL, and trajectory execution. It also provides grasping and pick-and-place pipelines.
MPC — see Model predictive control
Short for model predictive control.
Hub: Control.
MuJoCo
Multi-Joint dynamics with Contact, a general-purpose physics simulator maintained by Google DeepMind and released under the Apache 2.0 licence. It is widely used for robot learning and control research because it is fast and accurate with contact.
Multirotor
An aircraft lifted and steered by several fixed-pitch rotors, controlled by varying each rotor's speed. It can hover and manoeuvre precisely but has short endurance.

N

The ROS 2 navigation framework, which provides planners, controllers, costmaps, localisation and behaviour-tree-based task logic for mobile robots. It is led by Open Navigation LLC with community contributors.
Node
A unit of computation in ROS 2 that does one job, such as driving a sensor or running a planner, and communicates through topics, services and actions. Several nodes can share one process.
Nonholonomic constraint
A limit on velocity that cannot be integrated into a limit on position, such as a wheel that cannot slide sideways. A car can still reach any pose, but only by manoeuvring, so planners must respect the constraint.
Null space
The set of joint velocities that produce no end-effector motion, which exists for redundant arms. Controllers project secondary goals into it so they do not disturb the main task.

O

Object detection
Finding objects of known classes in an image and returning a box and a label for each. Robots usually need more than a box, so detection is followed by segmentation or pose estimation.
Object pose estimation
Estimating the 6-DoF pose of an object relative to the camera, so a robot can grasp or place it. Methods range from model-based matching to learned networks.
Observation space
The inputs a policy receives, such as joint states, camera images and force readings. Observations available only in simulation, such as exact object poses, must be removed or estimated before deployment.
Occupancy grid
A map that divides space into cells, each holding the probability that it is occupied. It is the standard map format for 2D navigation in ROS.
Odometry
Estimating how far a robot has moved by integrating its own motion measurements, such as wheel rotations or camera motion. Errors accumulate, so odometry drifts without external corrections.
OEE
Overall equipment effectiveness, the product of availability, performance and quality rates for a machine or line. It is a common measure for judging whether automation delivers in production.
Offline programming
Writing and testing robot programs in software against a model of the cell, then downloading them to the robot. It saves production time but needs good calibration because the real robot follows the numbers, not the model.
OMPL
The Open Motion Planning Library, an open-source collection of sampling-based planners from the Kavraki Lab at Rice University. It contains no collision checker of its own, so tools such as MoveIt 2 supply one.
Open-loop control
Sending commands without measuring the result, such as stepping a stepper motor a fixed number of steps. It works only while disturbances and model errors stay small.
Optical flow
The apparent motion of each pixel between two consecutive images. Drones use downward-facing flow sensors to hold position over the ground.
Overshoot
How far a step response passes its final value, given as a percentage of the step size. On a robot joint it shows up as the arm swinging past its target before settling.

P

Parallel manipulator
A mechanism in which several chains connect the base to one moving platform, as in a delta picker or a Stewart platform. Parallel designs are stiff and fast but have a smaller workspace than serial arms of similar size.
Parallel-jaw gripper
A gripper with two fingers that close towards each other along a straight line. It is simple and reliable and covers a large share of pick-and-place tasks.
Particle filter
An estimator that represents the belief as many weighted samples, moves them with the motion model and reweights them by how well they explain each measurement. It handles multimodal beliefs, such as several places a robot might be.
Path
A geometric curve through space or configuration space with no timing attached. A planner returns a path, and time parameterisation turns it into a trajectory.
Payload
The mass a robot can carry at its tool flange while meeting its rated performance, including the mass of the gripper itself. Makers usually rate payload at a stated distance of the load's centre of mass from the flange.
Peg-in-hole
The task of inserting a part into a tight-fitting hole, a standard test of compliant assembly. Small position errors cause jamming unless the robot complies to contact forces.
Performance level
The ISO 13849 measure of how reliably a safety function performs, graded from PL a, the lowest, to PL e, the highest. The risk assessment sets the required level for each safety function.
Physics engine
Software that steps rigid-body dynamics forward in time and resolves contacts and joint constraints. Different engines make different trade-offs between speed, stability and contact accuracy.
Pick and place
Grasping an object at one location and putting it down at another. It is the most common manipulation task in industry and a standard benchmark.
PID — see PID controller
Short for proportional-integral-derivative, as in PID controller.
Hub: Control.
PID controller
A feedback controller whose output is the sum of terms proportional to the error, its integral and its derivative. It is the default for joint position, velocity and current loops.
Pinhole camera model
The standard model of image formation in which each 3D point projects through a single centre onto the image plane. Real lenses add distortion on top of it.
PL — see Performance level
Short for performance level, as in PL d.
Planetary gearbox
A gearbox in which planet gears run between a central sun gear and an outer ring gear, sharing the load across several teeth. It is efficient and backdrivable at low ratios but has some backlash.
PLC
Programmable logic controller, a rugged industrial computer that runs machine logic in a fixed scan cycle. Robot cells are usually coordinated by a PLC, which talks to the robot controller over a fieldbus.
Pneumatic actuator
A cylinder or soft chamber driven by compressed air. It is cheap and naturally compliant, which suits grippers and suction tools, but precise position control is hard.
PoE — see Product of exponentials
Short for product of exponentials.
Hub: Kinematics.
Point cloud
A set of 3D points, sometimes with colour or intensity, as produced by lidars and depth cameras. ROS 2 carries them in sensor_msgs/PointCloud2 messages.
Policy
The function that maps what the robot observes to the action it takes. In robot learning it is usually a neural network.
Pose
The position and orientation of a frame relative to another frame. In 3D a pose has 6 degrees of freedom.
Pose graph
A graph whose nodes are robot poses and whose edges are measured relative motions between them, from odometry, scan matching or loop closures. Optimising it gives the most consistent trajectory.
Pose graph optimisation
Adjusting all poses in a pose graph at once to minimise the total disagreement with the measured edges, usually with nonlinear least squares. It is the back end of most SLAM systems.
Potential field
A planning method in which the goal attracts the robot and obstacles repel it, and the robot follows the combined gradient. It is fast but can trap the robot in local minima.
Power and force limiting
A collaborative method in which contact with a person may happen but the robot's speed, force and energy are limited so it stays below agreed thresholds. Most cobots are designed around it.
PPO — see Proximal policy optimisation
Short for proximal policy optimisation.
Prismatic joint
A joint that slides along a single axis without rotating. Linear axes on gantries and the vertical axis of a SCARA arm are prismatic.
PRM
Probabilistic roadmap, a sampling-based planner that scatters collision-free samples, links nearby ones into a graph and then searches the graph. The roadmap can be reused for many queries in the same environment.
Probabilistic completeness
The guarantee that a planner will find a path, if one exists, with probability approaching 1 as it runs longer. RRT and PRM are probabilistically complete but cannot prove that no path exists.
Process noise
The uncertainty a filter adds at each prediction step to account for an imperfect motion model. Setting it too low makes the filter overconfident and slow to follow real changes.
Product of exponentials
A way of writing forward kinematics as a product of matrix exponentials of screw axes, all expressed in one fixed frame. It needs no per-link frames, which makes it easier to set up than Denavit-Hartenberg.
Protective stop
A stop triggered automatically by a safeguard, such as a light curtain or a force limit, that halts the robot so it can resume once the cause is cleared. It is distinct from an operator's emergency stop.
Proximal policy optimisation
A widely used policy-gradient reinforcement learning algorithm that limits how far each update can move the policy. It is commonly used to train locomotion policies in parallel simulation.
Pure pursuit
A path-tracking method that steers towards a point on the path a fixed look-ahead distance ahead. A longer look-ahead gives smoother but less precise tracking.
PWM
Pulse-width modulation, which sets the average voltage applied to a load by switching it fully on and off at a fixed frequency and varying the on-time. Motor drives and hobby servos are both commanded this way.
PX4
Open-source flight control software for drones and other uncrewed vehicles, hosted by the Dronecode Foundation under the Linux Foundation. Its topics use north-east-down frames, while ROS 2 uses ENU, so bridges must convert.
PyBullet
The Python interface to the Bullet physics engine, used for robot simulation, learning and teaching. It can run headless in direct mode, which suits automated tests.

Q

QoS
Quality of service, the set of policies on a ROS 2 publisher or subscriber such as reliability, durability, history depth and deadline. A publisher and subscriber with incompatible QoS settings do not connect, which is a common source of silent failures.
Quadrature encoding
Two encoder signals, A and B, offset by a quarter of a cycle, so the order of their edges gives the direction of rotation. Counting every edge of both signals gives four counts per line.
Quadrotor
A multirotor with four rotors, two spinning each way so their reaction torques cancel. Yaw comes from speeding up one pair and slowing the other.
Quadruped
A four-legged robot. Quadrupeds are statically stable when standing, which makes them easier to deploy than bipeds, and are used for inspection and research.
Quasi-direct drive
An actuator that pairs a high-torque motor with a low-ratio gearbox, often a single planetary stage, so the joint stays backdrivable and transparent to impacts. Many quadrupeds use this design.
Quaternion
A four-number representation of rotation with no gimbal lock, where a unit quaternion encodes an axis and an angle. Libraries disagree on order: MuJoCo and spatialmath store (w, x, y, z), while ROS messages and SciPy use (x, y, z, w).

R

RaaS — see Robots as a service
Short for robots as a service.
Reach
The largest distance from the base axis to the wrist or tool flange that an arm can achieve, given in millimetres. Makers measure it from different reference points, so compare datasheets with care.
Real-time factor
The ratio of simulated time to wall-clock time. A value of 1 means the simulation keeps pace with reality, which matters when it is connected to real software or hardware.
Reality gap
The differences between a simulator and the real robot, in dynamics, contact, sensing, latency and appearance, that make a controller behave differently when moved to hardware.
Recursive Newton-Euler algorithm
An efficient inverse dynamics method that passes velocities outward from base to tip and forces inward from tip to base. Its cost grows linearly with the number of links.
Reflected inertia
The rotor inertia of a motor as seen at the output of a gearbox, which is multiplied by the square of the gear ratio. High ratios make the motor dominate the joint's inertia and reduce how safely the joint absorbs impacts.
Reinforcement learning
Learning a policy by trial and error, guided by a reward signal, rather than from labelled examples. In robotics it is mostly done in simulation because it needs millions of attempts.
Relocalisation
Recovering the robot's pose in an existing map after tracking is lost or at start-up. It uses place recognition and then a fine alignment.
REP 103
The ROS Enhancement Proposal that fixes units and coordinate conventions: SI units, right-handed frames with x forward, y left and z up for bodies, and ENU for local geographic frames. Camera frames with an _optical suffix use z forward, x right and y down.
REP 105
The ROS Enhancement Proposal that names the standard frames for mobile robots: map, odom and base_link. The odom frame is continuous but drifts, while map does not drift but can jump when localisation corrects it.
Repeatability
How closely a robot returns to the same taught position over many attempts, quoted as a ± value in millimetres. It is usually much better than absolute accuracy.
Revolute joint
A joint that rotates about a single fixed axis, like a hinge. Most robot arm joints are revolute, and their position is an angle in radians.
Reward function
The scalar signal that tells a reinforcement learning agent how well it is doing at each step. Poorly designed rewards are exploited in ways the designer did not intend.
RGB-D — see Depth camera
A colour image paired with a per-pixel depth image, as produced by a depth camera.
Hub: Sensors.
Right-hand rule
The convention that fixes axis directions: with the fingers of the right hand curling from x to y, the thumb points along z. The same rule gives the positive direction of a rotation about an axis.
Rigid-body dynamics
The study of how forces and torques produce motion in systems of bodies that do not deform. Robot arm and legged robot models nearly always assume rigid links.
Rise time
The time for a step response to go from a low to a high fraction of its final value, commonly 10 % to 90 %. Shorter rise time means a faster but often less damped loop.
Risk assessment
The documented process of identifying the hazards of a machine or application, estimating their risk and choosing measures to reduce it. Every robot installation needs one, and its findings set the required safety performance.
RL — see Reinforcement learning
Short for reinforcement learning.
RMW
The ROS middleware interface, the layer that lets ROS 2 run on different DDS implementations or other transports. You select one with the RMW_IMPLEMENTATION environment variable.
RNEA — see Recursive Newton-Euler algorithm
Short for the recursive Newton-Euler algorithm.
Hub: Dynamics.
Robot
A machine that senses its surroundings, decides what to do and acts on the physical world through motors or other actuators. The term covers fixed arms, wheeled and legged platforms, drones and underwater vehicles.
Robot cell
A self-contained work area holding one or more robots with their tooling, fixtures, conveyors and safeguarding. Cells are designed, guarded and certified as a unit.
Robot foundation model
A large model pre-trained on broad robot and internet data, intended to be adapted to many tasks and robots with little extra data. How well such models generalise is still an open research question.
Robots as a service
A business model in which the customer pays a subscription or per-task fee for robots instead of buying them, with the provider handling maintenance and updates. It is common for AMR fleets.
Hub: Careers and industry. Related: AMR
Roll, pitch and yaw
A common Euler angle convention that rotates about the x, y and z axes of a body, named after aircraft motions. URDF origins give orientation as roll, pitch and yaw in radians.
ROS — see ROS 2
Robot Operating System. On this site it means ROS 2 unless we say ROS 1.
ROS 2
Robot Operating System 2, an open-source set of libraries and tools for building robot software as communicating processes. It runs on top of an operating system rather than replacing one, and uses DDS for communication by default.
ROS bag
A recording of ROS 2 topic messages with timestamps, made with ros2 bag record and replayed with ros2 bag play. Storage plugins include MCAP and SQLite.
ROS distribution
A versioned, named set of ROS packages released together, such as Jazzy Jalisco or Lyrical Luth. A new ROS 2 distribution comes out each May, and those released in even years are long-term support releases with 5 years of support.
ROS package
The unit of organisation for ROS code, holding nodes, libraries, interface definitions, launch files and a package.xml that lists dependencies. Packages are built in a workspace with colcon.
ROS parameter
A named configuration value owned by a node, such as a gain or a frame name, that can be set at launch and changed at run time. Parameters are typically stored in YAML files.
ros2_control
The ROS 2 framework for real-time robot control, which separates hardware interfaces that talk to drives from controllers that compute commands. It is how most arms and mobile bases connect their hardware to ROS 2.
rosbag — see ROS bag
Another name for a ROS bag recording.
Rotation matrix
A 3 by 3 orthogonal matrix with determinant 1 that describes an orientation or rotates vectors between frames. Its columns are the axes of one frame written in the coordinates of the other.
ROV
Remotely operated vehicle, an underwater robot driven by a pilot through a tether that carries power and data. Offshore inspection and intervention rely heavily on ROVs.
RPY — see Roll, pitch and yaw
Short for roll, pitch and yaw.
RRT
Rapidly-exploring random tree, a sampling-based planner that grows a tree from the start by repeatedly extending it towards random samples. It finds a feasible path quickly in high-dimensional spaces, though the path is usually jagged.
RRT*
A variant of RRT that rewires the tree as it grows so the path cost keeps improving with more samples. It converges towards an optimal path but is slower per iteration.
RTK GNSS
Real-time kinematic positioning, which uses carrier-phase measurements and corrections from a nearby base station to bring GNSS error down to centimetre level. It needs a clear sky and degrades with distance from the base.
Hub: Sensors. Related: GNSS, Field robotics

S

Safety integrity level
The IEC 61508 family's discrete measure of the risk reduction a safety function must deliver, with higher numbers meaning lower allowed failure probability. A SIL belongs to a safety function, not to a whole machine.
Safety laser scanner
A 2D lidar certified for safety functions that triggers outputs when something enters configured protective and warning fields. AMRs and robot cells use them to detect people.
Safety-rated monitored stop
A collaborative method in which the robot stops and holds still under safety-rated monitoring while a person is in the shared space, then resumes automatically when they leave.
Sampling-based planning
Planning by drawing random configurations and connecting the collision-free ones, instead of building an explicit model of free space. It scales to arms with 6 or more joints where grid search becomes impractical.
Scan matching
Aligning a new lidar scan with a previous scan or a map to find the robot's motion. It gives much lower drift than wheel odometry in structured environments.
Screw axis
A line in space together with a pitch that describes a combined rotation about and translation along that line. Every rigid-body motion can be written as a motion along some screw axis.
SDFormat
Simulation Description Format, an XML format that describes robots, objects, environments and physics settings for simulators. It began as part of Gazebo and, unlike URDF, can describe whole worlds and closed chains.
SE(3) — see Special Euclidean group SE(3)
Notation for the special Euclidean group of rigid-body motions in 3D.
Semantic segmentation
Labelling every pixel of an image with a class, such as floor, wall or person. Instance segmentation also separates individual objects of the same class.
Sense-plan-act
The basic loop of a robot program: read the sensors, decide on an action, send commands to the actuators, and repeat. Real systems run several such loops at different rates, from a 1 kHz joint controller to a planner that updates a few times a second.
Sensor bias
A systematic offset in a sensor's output that noise averaging does not remove, such as a gyroscope reading a small rate when still. Bias often drifts with temperature, so estimators track it as part of the state.
Sensor fusion
Combining measurements from several sensors so that the estimate is better than any one of them alone. A typical mobile robot fuses wheel odometry, an IMU and occasionally GNSS or lidar localisation.
Sensor noise
Random variation in a measurement around its true value, often modelled as zero-mean Gaussian with a known variance. Filters weight each sensor by how noisy it is.
Serial manipulator
A robot arm built as one open chain, with each joint carrying all the joints after it. Most industrial arms and cobots are serial manipulators.
Series elastic actuator
An actuator with a deliberate spring between the gearbox and the load, so that measuring the spring's deflection gives the output force. The spring also absorbs shocks, at the cost of lower stiffness and bandwidth.
Service
A request-response call between ROS 2 nodes, used for short operations such as resetting a counter or querying a value. The caller waits for one reply, so services suit quick tasks only.
Servo motor
A motor run in closed-loop position, velocity or torque control using feedback from an encoder or potentiometer. In hobby robotics the word also names small all-in-one units commanded by a pulse width.
Settling time
The time after a step for the output to enter and stay within a band around its final value, commonly 2 % or 5 %. It is often what a cycle-time budget actually depends on.
Signed distance field
A map that stores, at each point, the distance to the nearest surface, negative inside objects. Planners and trajectory optimisers use it for fast collision costs and gradients.
SIL — see Safety integrity level
Short for safety integrity level. In simulation the same letters sometimes mean software-in-the-loop.
Sim-to-real transfer
Training or tuning a controller in simulation and then running it on the real robot. Success depends on closing or covering the reality gap through identification, randomisation or adaptation.
Simulation time step
The interval by which a simulator advances its state in one update, often a few milliseconds for robots with contact. Larger steps run faster but make contacts unstable.
Singularity
A configuration where the Jacobian loses rank, so the end-effector cannot move in some direction however the joints move. Near a singularity small tool motions demand very large joint speeds.
Skew-symmetric matrix
A 3 by 3 matrix equal to the negative of its transpose, built from a vector so that multiplying by it gives a cross product. Angular velocities and the matrix exponential of rotations are written with it.
Skid steer
Steering by driving the wheels or tracks on each side at different speeds, which forces them to slide sideways when turning. It is rugged and simple but makes wheel odometry poor.
SLAM
Simultaneous localisation and mapping: building a map of an unknown place while estimating the robot's pose within it. Most modern systems run a fast front end for odometry and a back end that optimises a pose graph.
SO(3) — see Special orthogonal group SO(3)
Notation for the special orthogonal group of 3D rotations.
Software-in-the-loop
Running the real control software, compiled for a desktop, against a simulator instead of the robot. PX4 and ArduPilot both offer it as SITL.
Special Euclidean group SE(3)
The set of all rigid-body motions in 3D, each a rotation combined with a translation. A pose and a homogeneous transformation are both members of SE(3).
Special orthogonal group SO(3)
The set of all 3D rotation matrices, which forms a group under matrix multiplication. Treating rotations as members of SO(3) rather than as loose numbers avoids many errors in filtering and optimisation.
Speed and separation monitoring
A collaborative method in which the robot slows or stops based on the measured distance to people, so contact never happens. It relies on safety-rated sensors such as laser scanners.
State estimation
Inferring quantities a robot cannot measure directly, such as its pose or velocity, from noisy sensor data and a motion model. The result is an estimate together with its uncertainty.
State-space model
A description of a system as first-order differential or difference equations in a state vector, with inputs and measured outputs. LQR, MPC and Kalman filters are all written in this form.
Static stability
Balance that holds at every instant without relying on motion, because the centre of mass stays above the support polygon. Slow walks with three feet down are statically stable, while trotting and running are not.
Steady-state error
The difference between setpoint and output that remains once the response has settled. An integral term removes it for constant setpoints and constant disturbances such as gravity.
Step response
How a closed-loop system's output moves after the setpoint jumps to a new value. Rise time, overshoot, settling time and steady-state error are all read from it.
Stepper motor
A motor that moves in fixed angular steps as its coils are energised in sequence. It can position without an encoder, but it loses steps silently if overloaded, so many robot uses add one.
Stereo camera
Two cameras a known distance apart whose images are matched to find depth from the disparity between them. Depth error grows with the square of distance and texture-less surfaces give no matches.
Strain wave gearing — see Harmonic drive (strain wave gear)
The generic name for the gear type sold under the Harmonic Drive trademark.
Support polygon
The convex hull of a robot's ground contact points. Static balance needs the centre of mass above it, and flat-foot dynamic balance needs the zero moment point inside it.
Synthetic data
Training or test data generated by simulation or rendering instead of recorded on real hardware. It comes with perfect labels but inherits the simulator's gaps.
System identification
Estimating the parameters of a model, such as masses, inertias and friction, from measured inputs and outputs of the real robot. Good identified models improve feedforward control and narrow the sim-to-real gap.
System integrator
A company that designs, builds and commissions complete robot applications from robots, tooling, safety equipment and software. Integrators carry the responsibility for the safety of the finished cell.

T

Tactile sensor
A sensor that measures contact over an area, such as a fingertip pad that reports pressure or the shape of what it touches. Camera-based designs image the deformation of a soft gel.
Task space
Describing motion in terms of the end-effector pose, usually in Cartesian coordinates. A straight-line tool move is planned in task space and converted to joint commands through inverse kinematics.
TCP — see Tool centre point
Short for tool centre point. In networking the same letters mean the transmission control protocol, so check the context.
Teach pendant
The handheld control unit of an industrial robot, used to jog the arm, teach positions and edit programs. It carries an emergency stop and an enabling switch.
Teleoperation
Controlling a robot remotely, with a person choosing its motions through a joystick, a leader arm, a VR headset or a similar device. Teleoperation is also the usual way to record demonstrations for imitation learning.
tf2
The ROS 2 library that keeps track of coordinate frames over time as a tree and answers questions such as where the camera was relative to the map at a given timestamp. Every pose in a ROS system names its frame for tf2.
Time parameterisation
Assigning timing to a geometric path so it respects velocity, acceleration and sometimes jerk limits. Time-optimal methods move as fast as the limits allow along the given path.
Time-of-flight camera
A depth camera that measures how long modulated light takes to return to each pixel. It works on texture-less surfaces but can be confused by sunlight, dark materials and multiple reflections.
Hub: Sensors. Related: Depth camera, Lidar
ToF — see Time-of-flight camera
Short for time of flight.
Hub: Sensors.
Tool centre point
The point on the end-effector whose position and orientation the controller tracks, such as the tip of a welding torch or the centre between gripper fingers. Setting it correctly in the controller is part of commissioning any new tool.
Tool changer
A coupling at the robot flange that lets the robot drop one tool and pick up another automatically, passing air and electrical signals through. It lets one arm do several jobs in a cell.
Topic
A named channel in ROS 2 over which nodes publish and subscribe to messages of one type. Publishers and subscribers do not know about each other, which suits streaming sensor data.
Torque constant
The torque a motor produces per ampere of current, in newton-metres per ampere. It lets a drive estimate output torque from measured current, though gearbox friction makes the estimate rough at the joint.
Trajectory
A path with timing, giving position, and usually velocity and acceleration, as a function of time. Controllers track trajectories, not paths.
Trajectory optimisation
Finding a trajectory by minimising a cost, such as time, effort or smoothness, subject to dynamics, limits and collision constraints. It produces smooth results but can get stuck in local minima.
Twist
A six-number vector holding the angular and linear velocity of a rigid body. ROS uses the same word for its velocity message, geometry_msgs/Twist, which mobile bases read as speed commands.

U

UAV
Uncrewed aerial vehicle, an aircraft with no pilot on board, flown remotely or autonomously. Drone is the everyday term for the same thing.
UKF — see Unscented Kalman filter
Short for unscented Kalman filter.
Ultrasonic sensor
A sensor that measures range from the echo of a sound pulse. It is cheap and sees glass, but it has a wide beam and coarse resolution.
Hub: Sensors. Related: Lidar, Sensor noise
Unicycle model
A simple kinematic model of a ground robot with a forward speed and a turn rate as inputs. Differential-drive robots map onto it directly.
Unscented Kalman filter
A Kalman filter for nonlinear systems that passes a small set of chosen sample points through the models instead of linearising them. It needs no Jacobians and often copes better with strong nonlinearity.
URDF
Unified Robot Description Format, the XML format ROS uses to describe a robot's links, joints, inertias, visual and collision geometry. It describes tree-shaped mechanisms only, so closed chains need workarounds.

V

Vacuum gripper
A gripper that holds objects with suction cups and a vacuum source. It suits flat, sealed surfaces such as boxes and sheets, and struggles with porous or uneven parts.
VIO — see Visual-inertial odometry
Short for visual-inertial odometry.
Vision-language-action model
A policy built on a large vision-language model that takes camera images and a text instruction and outputs robot actions. These models are trained on large multi-robot datasets and fine-tuned for a given robot.
Visual odometry
Estimating camera motion from how image features or pixel intensities move between frames. A single camera cannot recover absolute scale without other information.
Visual servoing
Controlling robot motion directly from camera feedback, by driving image features or an estimated pose towards a target. It tolerates calibration error better than look-then-move schemes.
Visual SLAM
SLAM that uses cameras as the main sensor, tracking image features or pixel intensities. It is cheap in hardware but sensitive to lighting, motion blur and texture-less scenes.
Visual-inertial odometry
Odometry that fuses camera images with IMU data, so the IMU gives scale and fast motion while the camera limits drift. It is the usual pose source for drones and headsets.
VLA — see Vision-language-action model
Short for vision-language-action model.
Voxel grid
A 3D occupancy grid made of small cubes. Octree versions store large empty regions cheaply, and voxel down-sampling is also a common way to thin a point cloud.
VTOL
Vertical take-off and landing. VTOL fixed-wing drones take off like a multirotor and transition to wing-borne flight for range.

W

Wheel odometry
Odometry from wheel encoder counts and the drive geometry. Wheel slip, uneven floors and inexact wheel radii make it drift, especially in heading.
Whole-body control
Computing all joint torques of a legged or mobile manipulator at once to meet several prioritised tasks, such as balance, foot contact and hand motion, under contact and torque limits. It is usually solved as a quadratic programme at each control step.
Workspace
The set of positions, or positions and orientations, that the end-effector can reach. The dexterous workspace is the smaller region where every orientation is possible.
Wrench
A six-number vector holding a force and a torque acting on a rigid body. Force-torque sensors report wrenches, and the Jacobian transpose maps a tool wrench to joint torques.

X

Xacro
An XML macro language used to write URDF files with variables, maths and reusable blocks. It is expanded into plain URDF before use.
Hub: ROS 2 and middleware. Related: URDF

Z

Zero moment point
The point on the ground where the horizontal moments from gravity and inertia forces cancel. A walking robot keeps its feet flat only while this point stays inside the support polygon.
ZMP — see Zero moment point
Short for zero moment point.