Glossary

Robotics has a lot of words, and most of them are used before they are explained. This page explains them.

It is meant to be looked up rather than read straight through. The terms are grouped by what they are about, and inside each group the simpler ones come first.

Where a term is explained properly somewhere else in this folder, there is a link.

Contents#

1. The arm itself
2. Describing where things are
3. Seeing
4. Moving
5. Touching
6. Deciding what to do
7. Learning
8. Measuring whether it worked
9. The frameworks

The arm itself#

Link. One rigid piece of the arm. The parts between the joints.

Joint. A place where two links move relative to each other. Most robot arm joints rotate. A few slide instead, and those are called prismatic joints.

Degree of freedom. One independent way something can move. A typical industrial arm has six joints, so it has six degrees of freedom. Six is the useful number because it is the smallest number that lets you put the gripper at any position and at any angle. With fewer, some angles become impossible. This is often shortened to DoF, as in "6-DoF".

End effector. Whatever is bolted to the end of the arm. A gripper, a welding torch, a suction cup, a screwdriver.

Gripper. The thing that holds objects. A parallel gripper has two fingers that close towards each other, and it is by far the most common kind. A suction gripper uses a vacuum cup, which is better for flat boxes and useless for porous objects.

Payload. How much weight the arm can carry at full reach. Be careful with this number: it usually includes the weight of the gripper, so the weight of the object you can actually lift is smaller.

Reach. How far the arm can extend. Usually measured from the base to the wrist.

Repeatability. How precisely the arm returns to a place it has been before. A good industrial arm manages about a tenth of a millimetre. This is the number vendors advertise.

Accuracy. How close the arm gets to a place it has been told about but never visited. This is a different number, and it is much worse than repeatability, usually by a factor of ten or more. Most people confuse the two. The gap between them is why calibration matters.

Singularity. An arm pose where the arm loses the ability to move in some direction, no matter what the joints do. The classic case is the arm stretched completely straight: it cannot extend any further. Near a singularity the maths demands enormous joint speeds for a small hand movement, so controllers either slow down or refuse. Planners avoid these poses.

Teach pendant. The handheld control box used to drive an industrial arm by hand and record positions. See teach and replay.

Lead-through, also called kinesthetic teaching. Programming the arm by physically pushing it where you want it to go, instead of driving it with buttons.

Collaborative robot, usually shortened to cobot. An arm built to work near people without a safety fence, normally by moving slowly and stopping when it feels an unexpected force.

Describing where things are#

Pose. A position and an orientation together. Where something is, and which way it is turned. A full pose in 3D has six numbers: three for position and three for rotation.

Frame. A set of axes that positions are measured against. Every robot system has many of them: one at the base of the arm, one at the gripper, one on each camera, one on the table. A position only means something when you say which frame it is in.

Transform. The relationship between two frames — how to convert a position in one into a position in the other. TF is the ROS system that keeps track of all of them and lets you ask "where is the gripper, in the camera's frame?" The arm area builds these up from scratch.

Joint space. Describing the arm by its joint angles. Six numbers for a six-joint arm. This is what the motors actually care about.

Cartesian space, also called task space. Describing the arm by where the gripper is, in ordinary x, y and z. This is what you care about.

Forward kinematics. Given the joint angles, work out where the gripper is. This is the easy direction: there is exactly one answer, and it is a chain of multiplications.

Inverse kinematics, usually shortened to IK. Given where you want the gripper, work out the joint angles. This is the hard direction. There may be several answers — an arm can often reach the same place with its elbow up or down — and sometimes there are none, because the place is out of reach.

Workspace. All the poses the arm can actually reach. It is a strange shape, and it has holes in it, usually directly above and behind the base.

URDF. The Unified Robot Description Format, an XML file that lists a robot's links, joints, shapes and limits. ROS, Gazebo and MoveIt all read it.

MJCF. MuJoCo's own robot description format. Same job as a URDF, different file. The same arm needs both if you use both simulators, and keeping them in agreement is a real chore.

Seeing#

Depth camera. A camera whose pixels record distance rather than colour. An RGB-D camera records both, so you get a colour picture and a distance for each pixel.

Point cloud. A set of 3D points, usually produced from a depth picture. This is what most robot perception code actually works on, because points can be reasoned about geometrically while pixels cannot.

Segmentation. Working out which pixels belong to which object. Doing this well for objects the system has never seen is what SAM 2 provides, and it is the main reason machine learning has been adopted in real robot cells.

Pose estimation. Working out the full position and orientation of a known object. Often called 6-DoF pose estimation, because the answer has six numbers.

Occlusion. One object hiding another from the camera. This is the central difficulty of bin picking, because in a pile most objects are partly hidden.

Calibration. Measuring the fixed relationships that your model assumes. Hand-eye calibration is the important one: it works out exactly where the camera is relative to the arm. If this is wrong by two millimetres, then everything the camera reports is wrong by two millimetres, and no amount of good software fixes it. Calibration is unglamorous and it decides whether a real cell works.

Grasp pose. Where the gripper should be, and how it should be turned, in order to pick something up. A grasp proposal system suggests several and ranks them.

Antipodal grasp. A grasp where the two fingers close on opposite, roughly parallel surfaces. This is the shape of grasp a parallel gripper wants, and finding these on a point cloud is the basis of a simple hand-written grasp generator.

Moving#

Path. A sequence of positions from here to there. It says nothing about speed.

Trajectory. A path with timing attached — where the arm should be, and when. Controllers execute trajectories, not paths.

Waypoint. One recorded position along a path.

Motion planning. Finding a path from the arm's current pose to a target pose without hitting anything. Explained fully in the programmed methods.

Configuration space, often shortened to C-space. The space of all possible joint angle combinations. A six-joint arm has a six-dimensional configuration space. Planners search in this space rather than in ordinary 3D space, and the reason planning is hard is that obstacles become complicated, unpredictable shapes once you translate them into it.

Sampling-based planner. A planner that throws random configurations at the problem, throws away the ones that collide, and connects the rest until a path appears. RRT, the rapidly-exploring random tree, and PRM, the probabilistic roadmap, are the two standard ones. They are good at finding a way through awkward spaces. They give a different answer every run, which is a genuine problem if anybody has to certify what the machine will do.

Optimisation-based planner. A planner that starts with a guess at the whole path and pushes it away from obstacles while keeping it short and smooth. CHOMP and TrajOpt are the classic ones. The paths are nicer and repeatable, but the method can get stuck where a random planner would eventually have found a way.

Point-to-point, linear and circular motion. The three motion types an industrial controller actually offers. Point-to-point moves the joints directly, linear moves the gripper in a straight line, circular moves it along an arc. They are deterministic, which is why factories use them.

Collision checking. Testing whether a given arm pose overlaps anything, including the arm overlapping itself. Planners do this thousands of times per query, so it has to be fast.

Jerk. The rate at which acceleration changes. Limiting it is what stops an arm from moving in a way that shakes the whole cell. Ruckig does this.

Touching#

Position control. Telling the arm where to be. The controller then pushes as hard as it needs to in order to get there. This is the default, and it is what breaks parts: commanded into a surface, it sees an error it cannot remove and pushes harder.

Force control. Telling the arm how hard to push, rather than where to be.

Compliance. How much the arm gives way when something pushes it. A compliant arm is soft; a stiff arm is not.

Stiffness. The number that says how compliant the arm is — how much force it produces for a given amount of displacement. Choosing this number deliberately, rather than inheriting whatever the mechanics give you, is the whole point of impedance control.

Impedance control. Making the arm behave like a spring of a stiffness you chose. Push it and it pushes back gently, in proportion to how far you moved it.

Admittance control. The same idea approached from the other side: measure the force, then move in response to it. Which of the two you use depends mostly on what your hardware can sense. Both are covered in feedback control.

Force-torque sensor. A sensor, usually at the wrist, that measures the forces and twists acting on the gripper. Accurate and expensive, and a lot of research deliberately avoids needing one.

Compliant flange. A small spring-loaded device bolted between the arm and the tool, which holds a set contact force over a few millimetres of travel. It reacts far faster than the arm can, which is why polishing and deburring use one. It is a mechanical answer to a control problem.

Visual servoing. Closing the control loop on the camera instead of on the joint sensors. You move so as to reduce the difference between what the camera sees and what it should see. Its virtue is that it never needs to know where anything is in world coordinates.

Model predictive control, usually shortened to MPC. Repeatedly solving a short planning problem: given where I am now and a model of how things move, what is the best sequence of commands over the next second? It executes only the first command, then throws the rest away and solves again. Re-solving is how it absorbs everything the model got wrong.

Deciding what to do#

State machine. A set of named states with rules for moving between them. Easy to follow with five states, unreadable with fifty, because the number of possible transitions grows with the square of the number of states.

Behaviour tree. A task arranged as a tree of steps, run repeatedly. Each step reports success, failure, or still running. A sequence node runs its children in order until one fails. A fallback node tries each child until one succeeds, which is the shape of "try the normal thing, and if that fails, try the recovery". They stay readable at fifty branches, which is why they won.

Primitive, or skill. One named, reusable piece of robot behaviour, such as "approach", "grasp" or "retreat". Real systems are built by sequencing these.

Task and motion planning, usually shortened to TAMP. Searching for what to do and how to move at the same time, because sometimes they cannot be separated. Very capable, very slow, and used far less than its reputation suggests.

Learning#

Policy. A function that takes what the robot can see and produces what the robot should do next. When people say "the model" in robot learning, they usually mean this.

Demonstration. One recorded example of a person doing the task, usually by driving the robot. The recording includes the camera pictures, the joint angles and the commands the person gave.

Teleoperation. Driving the robot remotely so that its motion can be recorded. For arms this is often done with a leader arm, which is a small copy of the robot that you move with your hand while the real arm follows.

Imitation learning, also called learning from demonstration. Training a policy to copy what the demonstrator did.

Behaviour cloning. The simplest form of imitation learning: train the network to output the command the person gave, given what the cameras saw. Explained in the learned methods.

Compounding error. The central problem of behaviour cloning. A slightly wrong command puts the arm somewhere slightly unlike the training data, where the next command is a little worse, and within a second or two the arm is somewhere no demonstrator ever went. Errors build with every decision taken.

Action chunking. Predicting the next hundred commands at once and playing them out, rather than predicting one command at a time. If you make a hundred times fewer decisions, you get a hundred times fewer chances to drift. This is the main fix for compounding error.

ACT. Action Chunking with Transformers. The specific policy that made action chunking popular, and still the recommended first thing to try for a new task. It trains in under an hour on one ordinary graphics card, which is most of why it is recommended.

Diffusion policy. A policy that generates a whole action sequence the way image models generate pictures, by starting from noise and refining. The reason to bother is multimodality: if some demonstrators went left around an obstacle and others went right, a plain network averages them and drives straight into it. A diffusion policy can represent "either this or that".

Flow matching. A faster relative of diffusion that solves the same problem. In 2026 it is the action-generating part inside nearly every large policy. Worth knowing that a 2026 study found no measurable advantage over plain regression on its benchmarks, so this is not as settled as its popularity suggests.

Multimodal. Having several valid answers rather than one. The word is used this way in robot learning, which is unrelated to its other meaning of "using several kinds of input".

Reinforcement learning, usually shortened to RL. Learning by trying things and keeping what worked, guided by a reward rather than by demonstrations.

Reward. A number that says how well things are going. Writing one is harder than it sounds, because any gap between what you wrote and what you meant will be found and exploited.

Model-free and model-based RL. Model-free methods learn directly from experience with no model of the world, which is simple and needs enormous amounts of practice. Model-based methods learn a model first and then practise inside it, which needs far fewer real attempts and has many more moving parts.

Offline RL. Learning from a fixed pile of recorded behaviour, with no practising at all.

DAgger. Dataset Aggregation. The idea of letting the policy drive, correcting it when it goes wrong, and adding those corrections to the training data. Corrections cover the situations the policy actually reaches, which is why a few dozen of them beat several hundred fresh demonstrations.

HIL-SERL. Human-in-the-Loop Sample-Efficient Robot Learning. The modern version of that idea, where your take-overs become the reinforcement learning signal. It has the strongest published numbers in this folder.

Sim-to-real gap. The difference between the simulator and the world, and the reason a policy that works in simulation may not work on hardware.

Domain randomisation. The main fix for that gap. Instead of building one accurate simulator, you train across thousands of randomised ones — varying friction, mass, lighting and delays — so that reality is just one more variation the policy has already coped with.

Foundation model. A large model trained on a great deal of general data, meant to be adapted to specific jobs rather than used as-is.

Pretraining and fine-tuning. Pretraining is the expensive first stage on a large general dataset. Fine-tuning is the cheap second stage that adapts the result to your task with comparatively few examples. On this path you fine-tune and never pretrain.

Vision-language-action model, usually shortened to VLA. A large pretrained policy that takes camera pictures and a sentence, and produces arm commands. The sentence is what lets one model do many tasks.

Checkpoint, or weights. The saved, trained state of a model — the actual file you download. The distinction that matters in practice is between a model that has been announced and one whose checkpoint you can actually download. Several of the most impressive models of 2026 have no released weights.

World model. A model that learns to predict what will happen next, so the policy can act through that prediction. The interesting property is that it can learn from video with no recorded actions attached, which matters because video is plentiful and robot demonstrations are not.

Retargeting. Converting a motion recorded from one body onto another — most often a human hand onto a robot gripper. It is the hard, unglamorous core of learning from human video, because the two have different shapes and different joints.

Measuring whether it worked#

Episode, or rollout, or trial. One attempt at the task, from the starting state to success or failure. These three words mean the same thing and different papers pick different ones.

Success rate. The fraction of attempts that worked. Meaningless without the number of attempts.

Denominator. The number of trials behind a percentage. A 90% success rate over ten trials and over a thousand trials are completely different claims. If a source gives a percentage and no denominator anywhere, treat it as marketing. This habit is worth more than any single technique, and the overview explains why.

Ablation. An experiment that removes one piece of a system to find out how much that piece was contributing. The most informative kind of result, and the one most often missing.

Benchmark. A fixed set of tasks everyone tests on, so that results can be compared. Useful, and worth remembering that a method tuned on a benchmark is partly tuned to that benchmark.

The frameworks#

Short entries. The learning path explains each of these in depth, including why to choose it over the alternative.

The "Mac" column says whether it runs on an Apple Silicon Mac without an NVIDIA graphics card.

Simulators#

Motion and control#

Sequencing#

Seeing#

Learning#

Back to the overview, or to the learning path if you want to build something.