Our Work — Glossary
Robots and AI, in plain language.
153 terms and questions, written for people who know nothing about the field and still accurate for people who do. Search, or filter by type, level, and topic.
Questions
36
- QuestionBeginner
Are robots going to take my job?
Rarely the whole job, and never overnight. Robots follow the same law of scaling as every tool before them: the tractor did not abolish farming, it let one farmer cultivate far more land. Robots will enter the workforce gradually, first amplifying existing practices, then requiring progressively less human input as their capabilities mature. Because the shift is incremental, people have time to adapt their skills and careers. The result is greater output per person. AI paired with a physical presence allows fewer hands to accomplish ever more, expanding what the world can produce. We call this “doing more with the same people,” and advocate for fair retraining along the way.
- QuestionBeginner
Are robots safe to be around?
Treat them as machines, because that is what they are. Until deployment is widespread and regulation has matured, extend them the same caution you would any powerful equipment: a robot can injure a person, and children are especially vulnerable. There is no need to be scared if you encounter one in public. Simply stay aware, keep a sensible distance, and be ready should it move suddenly.
- QuestionBeginner
Can a robot understand me?
Not the way a person does. A robot can turn your words into text, work out what you probably want, and act on it. It has no feelings and no real grasp of meaning, so it can misread tone or intent. Honest design does not pretend otherwise.
- QuestionBeginner
Can robots have emotions or feelings?
No. A robot can show a smile or a friendly voice, but these are programmed or learned performances, not real feelings. They make interaction easier, and they can also manipulate people, especially children. Clear labeling helps people remember it is a machine.
- QuestionBeginner
How do I talk to or work alongside a robot safely?
Treat it like moving machinery. Stay aware of its work area, avoid surprising it from behind, and know where the emergency stop is. Give clear commands if it takes voice input. Follow floor markings and site rules, and if it acts oddly, stop it and report it.
- QuestionBeginner
How do robots learn?
Mostly by example and by practice. In imitation learning, a person shows the task, often by remote-controlling the robot, and it copies the pattern. In reinforcement learning, the robot tries many times, usually in simulation, and is rewarded when it gets closer. Many robots combine both with large pretrained models.
- QuestionBeginner
How do robots see?
With sensors and software instead of eyes and a brain. Cameras capture images, while depth cameras and LiDAR measure distance. Perception software turns that data into simple facts, like “there is a cup 30 centimeters ahead,” which the robot then acts on.
- QuestionIntermediate
How much energy do robots and AI use?
The big cost is the data centers that train and run AI models, not the robots themselves. A robot runs on a battery like a power tool or small vehicle. Gartner expects data centers to use 565 terawatt hours in 2026, up from 447 in 2025, so power supply now limits where AI can grow.
- QuestionBeginner
Should robots look like humans?
Not always. A human shape helps a robot fit spaces and tools built for people. But it can also make people over-trust the machine, or feel uneasy when it looks almost human. Many jobs are better done by a simple arm, cart, or drone. Form should follow function and honesty.
- QuestionBeginner
What are drones?
Drones are aircraft that fly without a pilot on board, controlled remotely or by software. Small ones use spinning rotors and carry cameras for filming, inspection, farming, and delivery. Because they fly over people and property, airspace and privacy rules govern where they can go.
- QuestionIntermediate
What are robots made of?
Mostly metals like aluminum, steel, and titanium for the frame, copper for wiring, plus carbon fiber, plastics, and silicone. The most important materials are rare earth elements, especially neodymium and dysprosium, which make the strong magnets inside electric motors. China did about 60 percent of the mining and 91 percent of the refining in 2024, and its recent export controls caused shortages and higher prices. Robots and electric cars need the same magnets, so supply is a real policy issue.
- QuestionBeginner
What are the components of robots?
Most robots share the same building blocks. Sensors take in the world, like cameras and force sensors. Motors and other actuators create movement, with gearboxes adjusting the force. A controller and onboard computer make decisions, powered by a battery or a cable. An end effector, like a gripper or tool, does the work, all mounted on a frame. Factory arms, warehouse robots, drones, surgical robots, and humanoids mix these same parts in different ways.
- QuestionAdvanced
What are the different kinds of models that power robots?
Robots use different AI than chatbots because they must produce movement, not text. Vision-Language-Action models (VLAs) take camera images and an instruction and output motor actions. World models predict how a scene will change. World action models (WAMs) also output the moves to cause it. ACT is an imitation method that predicts short chunks of actions so motion stays smooth. Companies often start from robot foundation models like pi0, Gemini Robotics, or NVIDIA GR00T and fine-tune them.
- QuestionBeginner
What does the Institute mean by humane robotics?
Robots that respect people's safety, dignity, consent, and privacy, so daily contact with them feels natural rather than stressful. We work toward this in three ways: policy, consulting with AI and robotics companies, and public education. Our phrase “doing more with the same people” means robots should expand what people can do, not shrink the workforce.
- QuestionIntermediate
What happens to a robot's data and my privacy?
Robots with cameras and microphones can record your face, voice, home, and habits. Some keep that data on the device, others send it to the cloud, and policies vary. Good practice is to collect the least data needed, store it securely, and let you see and delete it. Ask before you buy or work with one.
- QuestionBeginner
What is a chatbot or LLM?
A chatbot is a program you talk to in plain language. Most run on a large language model, or LLM, an AI trained on huge amounts of text to predict the next word and write fluent replies. Common ones in 2026 include ChatGPT (OpenAI), Gemini (Google), Claude (Anthropic), Microsoft Copilot, Grok (xAI), Meta's Llama, and DeepSeek, with ChatGPT the most used at about 54 percent of visits. They are good at writing and summarizing but can be confidently wrong, and a chatbot is not a robot.
- QuestionBeginner
What is a cobot?
A cobot, short for collaborative robot, works in the same space as people without a safety cage. It moves slowly and gently enough that contact does not injure. Cobots do jobs like assembly and packaging next to workers, and safety guidance like ISO/TS 15066 sets their limits.
- QuestionBeginner
What is a companion robot?
A companion robot keeps people company, gives reminders, or offers a sense of presence, often for older adults or people living alone. It can ease loneliness and help with routines. It also raises questions about emotional attachment, honesty, and the personal data it collects.
- QuestionBeginner
What is a humanoid robot?
A humanoid robot has a human-like body, usually a torso, arms, legs, and a head, so it can work in spaces built for people. Balancing on two legs and using hands are very hard. In 2026 most humanoids are in factory and warehouse pilots, not homes.
- QuestionBeginner
What is a robot?
A robot is a machine that senses its surroundings, decides what to do, and physically acts, with some independence. That sense-think-act loop separates it from a toaster, which just runs a fixed cycle. Robots range from factory arms and warehouse carts to drones, surgical systems, and walking humanoids.
- QuestionIntermediate
What is a self-driving car and is it a robot?
Yes, it is a robot on wheels. The industry rates cars from Level 0 (no automation) to Level 5 (drives itself anywhere). Cars you can buy today top out at Level 2, so you are still the driver. Only limited robotaxi services like Waymo run at Level 4. Marketing terms often overstate this.
- QuestionIntermediate
What is a surgical robot?
A surgical robot lets a surgeon operate through tiny cuts with steadier, finer movements than a hand alone. The surgeon controls it in real time, so it is not acting on its own. Benefits include smaller incisions and faster recovery. Because a person stays in control, responsibility stays clear.
- QuestionAdvanced
What is a world model?
A world model is an AI that learns how a scene changes over time, so it can predict what happens next. Think of it as an internal simulator of the physical world. Examples include NVIDIA Cosmos and Google DeepMind's Genie 3. Robots can use one to plan and practice before acting.
- QuestionBeginner
What is AI?
AI is software that does tasks we usually link with human thinking, like recognizing images, understanding language, or making decisions. Most modern AI learns patterns from lots of examples instead of following hand-written rules. It is powerful but narrow: good at what it was trained on, and prone to failing outside that.
- QuestionIntermediate
What is embodied AI?
Embodied AI is intelligence inside a physical body, like a robot, that learns by sensing and acting in the real world. It is much harder than a chatbot because the world does not pause and small errors add up fast. It is the core of modern robotics research.
- QuestionBeginner
What is human-robot interaction?
Human-robot interaction, or HRI, is the study of how people and robots communicate, share space, and work together. It covers practical things like how a robot signals its next move, and social things like trust and comfort. Good HRI makes robots predictable and easy to read. It is the heart of the Institute's work.
- QuestionIntermediate
What is simulation and why do robot companies use it?
Simulation is a virtual copy of the world where a robot can practice safely, quickly, and cheaply. A robot can try a task millions of times overnight without wearing out or hurting anyone. The catch is the sim-to-real gap: skills that work in the clean virtual world can fail in messy reality.
- QuestionIntermediate
What is teleoperation?
Teleoperation means a person controls a robot remotely in real time, like a surgeon guiding instruments or an operator driving a warehouse robot from a screen. It is used when a task is too hard for autonomy or too risky for a person. Recorded teleoperation is also a main source of robot training data.
- QuestionBeginner
What is the difference between automation and AI?
Automation follows fixed rules and does a task the same way every time, like a dishwasher cycle. AI learns patterns from data and can handle situations it was never explicitly programmed for. Many useful robots are pure automation, which is often the safest choice. AI adds flexibility but also unpredictability.
- QuestionIntermediate
What is the difference between LLMs and robots?
An LLM handles language, while a robot is a physical machine that senses and acts in the real world. An LLM can help power a robot, for example by turning your request into a plan, but the robot still needs its own systems for seeing, moving, and staying safe. Chatbot mistakes are annoying. Humanoid mistakes can hurt someone, so robots need their own rules covering physical safety, consent before touching or recording you, dignity, privacy, and predictable behavior around people.
- QuestionAdvanced
What is the EU AI Act and does it cover robots?
The EU AI Act is Europe's law for AI, in force since 2024, which sorts systems by risk and puts the strictest duties on high-risk uses. Robots are covered mainly through the separate Machinery Regulation, which applies from January 2027. So an AI robot in Europe faces rules for the AI, the machine, and any personal data it collects.
- QuestionBeginner
What is the uncanny valley?
The uncanny valley is the uneasy feeling people get from something that looks almost human but slightly off, like a lifelike face with stiff movement. Comfort rises as a design gets more human, then drops sharply when small flaws feel wrong. Designers often choose a friendly, clearly machine look to avoid it.
- QuestionAdvanced
What laws or rules exist for robots today?
There is no single robot law, but many rules already apply. ISO 10218 covers industrial robots, ISO/TS 15066 covers cobots, and ISO 13482 covers some personal-care robots. Europe adds the AI Act and Machinery Regulation, China issued its first national humanoid standards in 2026, and privacy laws apply whenever a robot records people. Rules for humanoids in public and homes still have gaps.
- QuestionAdvanced
Who is responsible when a robot causes harm?
Usually a person or company, not the robot. It could be the maker if the design was faulty, the employer or operator if it was misused, or the installer who set it up. AI that behaves unpredictably makes blame harder to assign. Clear records and human oversight keep responsibility traceable.
- QuestionIntermediate
Why do robots fail at simple tasks humans find easy?
Because “easy” for people is not easy at all. Walking on gravel or folding a towel uses skills evolution built into us over millions of years, so they feel effortless. A robot has to rebuild all of that from sensors and code. This is Moravec's paradox: math is easy for machines, movement is hard.
- QuestionIntermediate
Why do robots need so much data?
Because they learn from examples, and the more good examples they see, the better they handle new situations. Text for chatbots exists in vast amounts online, but robot data must be collected action by action, often by a person remote-controlling the robot. That is slow and expensive, so companies also use simulation and shared datasets.
Terms
117
- TermAdvanced
ACT (Action Chunking with Transformers)
A robot learning method that predicts a short chunk of future actions at once, which keeps motion smooth. It comes from the ALOHA project and learns by copying human demonstrations.
- TermBeginner
Actuator
The part of a robot that creates movement, like a muscle. Common types are electric motors, hydraulic cylinders, and pneumatic pistons.
- TermBeginner
AGV (automated guided vehicle)
A wheeled robot that moves materials along fixed routes, following wires, magnets, or markers in the floor. Reliable but not flexible.
- TermIntermediate
Alignment
The work of making an AI's behavior match what people actually want, so it helps and avoids harm. It gets harder as systems become more capable.
- TermBeginner
AMR (autonomous mobile robot)
A mobile robot that navigates on its own using sensors and maps, avoiding obstacles rather than following fixed tracks. Common in warehouses and hospitals.
- TermBeginner
Anthropomorphism
The human habit of seeing feelings and intentions in non-human things, including robots. It makes interaction feel friendly but can mislead people about what a machine really is.
- TermIntermediate
Automation bias
The tendency to over-trust automated systems and stop checking their work. It can lead someone to accept a bad robot decision because “the machine said so.”
- TermIntermediate
Autonomy levels
A scale describing how much a machine does on its own versus how much a human controls. The best-known example is the 0 to 5 scale for self-driving cars.
- TermBeginner
Battery / energy density
A battery stores the energy a mobile robot runs on. Energy density is how much energy fits in a given weight, which limits how long legged robots can work.
- TermIntermediate
Behavior cloning
The simplest form of imitation learning, where a robot copies recorded human actions directly. It is easy to set up but fails in situations it never saw.
- TermIntermediate
Benchmark
A standard test used to compare AI models or robots on the same task. Useful for measuring progress, but a model can be tuned to the test rather than reality.
- TermIntermediate
Bias
When an AI produces skewed or unfair results, usually because its training data was unbalanced. In robots, this could mean recognizing some people less reliably than others.
- TermIntermediate
Brushless DC motor
An efficient, long-lasting electric motor with no physical brushes to wear out. It is common in drones and robot joints and usually relies on rare earth magnets.
- TermBeginner
Camera / computer vision
A camera captures images, and computer vision is the software that makes sense of them. It is the most common way robots see, but struggles in poor light.
- TermBeginner
Care robot
A robot that helps with health or daily tasks, like helping someone stand or reminding them to take medicine. It needs high standards for safety, consent, and privacy.
- TermBeginner
Chassis
The main frame of a robot that everything else attaches to. It is usually made of aluminum, steel, or carbon fiber, balancing weight against strength.
- TermBeginner
Cobot
A collaborative robot built to work next to people without a cage, moving gently enough that contact does not injure. Common for assembly and packaging.
- TermAdvanced
Compliant control / impedance control
Control that lets a robot give a little when it meets resistance, like a spring, instead of pushing rigidly. Essential for safe contact with people and objects.
- TermIntermediate
Consent (in human-robot interaction)
A person knowingly agreeing before a robot touches them, records them, or acts for them. It should be clear and ongoing, not buried in fine print.
- TermIntermediate
Context window
How much text an AI language model can consider at once, like its short-term memory. Once it fills up, older content drops out of view.
- TermIntermediate
Controller / control loop
The controller decides how a robot moves. The control loop is the constant cycle of sensing, comparing to the goal, and correcting, repeated many times per second.
- TermIntermediate
Cybersecurity for robots
Protecting robots from hacking and tampering. A hacked robot is not just a data risk, because an attacker could take over how it moves.
- TermIntermediate
Deep learning
A type of machine learning that uses neural networks with many layers to find complex patterns in large amounts of data. It powers most modern AI.
- TermIntermediate
Degrees of freedom
The number of independent ways a robot can move. A simple gripper may have one, while a humanoid can have dozens. More freedom means more flexibility and more complexity.
- TermBeginner
Deployment
Putting a robot or AI to work in the real world, outside the lab. It is where hidden problems show up, so safety decisions must be settled first.
- TermBeginner
Depth camera
A camera that measures how far away things are, giving the robot a 3D view. It helps with grasping and avoiding obstacles.
- TermIntermediate
Dexterous hand
A robot hand with multiple fingers and joints, built to handle objects with human-like skill. Powerful but complex, costly, and hard to control.
- TermAdvanced
Diffusion policy
A method that generates robot actions the way image generators build pictures from noise, refining a rough plan into smooth motion. Good at tasks with several valid solutions.
- TermIntermediate
Digital twin
A detailed virtual copy of a real robot, machine, or factory, kept in sync with the real thing. Used to test changes and predict problems safely.
- TermAdvanced
Domain randomization
A training trick where a robot practices in simulation with wildly varied lighting, colors, and physics, so the real world looks like just another variation.
- TermBeginner
Drone / UAV
An aircraft that flies without a pilot on board. Drones carry cameras and sensors for filming, inspection, and delivery, and are governed by airspace and privacy rules.
- TermIntermediate
Edge computing
Processing data on or near the robot instead of in a distant data center. It cuts delay, works without a connection, and keeps sensitive data local.
- TermIntermediate
Embodied AI
Intelligence inside a physical body that senses and acts in the real world, unlike software that only handles text or images. It must cope with real physics and timing.
- TermIntermediate
Encoder
A sensor that measures the position or rotation of a motor or joint, telling the robot exactly where its parts are. Without it, movement would be blind.
- TermBeginner
End effector / gripper
The tool at the end of a robot arm that does the actual work, such as a gripper, suction cup, or welder. It is the robot's working hand.
- TermBeginner
Exoskeleton
A wearable robotic frame that supports or boosts a person's own movement, like helping a worker lift or a patient walk. The person stays in control.
- TermIntermediate
Explainability
How well people can understand why an AI or robot made a decision. Many modern models are black boxes, which is a problem when something goes wrong.
- TermBeginner
Fail-safe / emergency stop
A fail-safe moves a robot to a safe state when something goes wrong. An emergency stop is the button that halts it immediately. Both are required near people.
- TermIntermediate
Fine-tuning
Training an already-trained model a bit more on specific data for a particular job. It is how companies adapt a general robot model to their own robot.
- TermIntermediate
Fleet management
The software and practice of running many robots at once, tracking their tasks, charging, updates, and health. Essential once robots are deployed at scale.
- TermIntermediate
Force/torque sensor
A sensor that measures pushing and twisting forces, usually at a robot's wrist. It gives the robot a sense of feel for delicate tasks and safe contact.
- TermAdvanced
Foundation model
A large AI model trained on broad data that can be adapted to many tasks. LLMs are one kind; robot versions include RT-2, pi0, and NVIDIA GR00T.
- TermAdvanced
Functional safety
The engineering discipline of making sure safety features keep working even when parts fail, using redundancy and testing. It underpins emergency stops and safe speed limits.
- TermAdvanced
Generalist robot policy
A single AI model that can perform many different robot tasks, rather than one model per task. It is the goal of most robot foundation models.
- TermBeginner
GPU
A chip that does many calculations at once, which makes it ideal for training and running AI. GPUs power both data centers and some robots' onboard brains.
- TermIntermediate
Guardrails
Built-in limits that stop an AI or robot from doing unsafe or unwanted things, like capping a robot's speed or force. A safety layer on top of the model.
- TermBeginner
Hallucination
When an AI states something false as if it were true, because it predicts plausible words rather than checking facts. A robot acting on one could make a wrong move.
- TermAdvanced
Harmonic drive / gearbox
A gearbox trades motor speed for force. A harmonic drive is a compact, very precise gearbox popular in robot joints because it has almost no slack.
- TermIntermediate
Human in the loop
A person is directly involved in a system's decisions, approving or correcting actions before they happen. Slower, but safer and easier to hold accountable.
- TermIntermediate
Human on the loop
A person supervises an autonomous system and can step in, but does not approve every action. A middle ground between full autonomy and human in the loop.
- TermBeginner
Humanoid robot
A robot with a broadly human body shape so it can work in human spaces and use human tools. Balance and hands make it especially hard to build.
- TermAdvanced
Hydraulic vs electric actuation
Hydraulic joints use pressurized fluid and deliver huge force but are heavy and leak-prone. Electric motors are cleaner and more precise, and now dominate modern humanoids and cobots.
- TermIntermediate
Imitation learning
Teaching a robot by showing it examples of a task, usually recorded from a person, so it copies what it saw. Fast, but weak in unfamiliar situations.
- TermIntermediate
IMU (inertial measurement unit)
A small sensor that measures motion and orientation, telling a robot how it is tilting, turning, and accelerating. Essential for balance in legged and flying robots.
- TermBeginner
Industrial robot arm
A jointed mechanical arm used in factories for welding, painting, and assembly. Fast and precise, but usually caged because it cannot sense people.
- TermIntermediate
Inference
Using a trained AI model to produce an answer or action, as opposed to training it. It happens every time a chatbot answers or a robot decides its next move.
- TermAdvanced
Inverse kinematics
The math a robot uses to work out which joint angles will put its hand or tool at a desired spot. It runs constantly behind every reach or step.
- TermAdvanced
ISO 10218
The main international safety standard for industrial robots and their installation. Updated in 2025 to focus more on how robots are used, not just built.
- TermAdvanced
ISO 13482
An international safety standard for certain personal-care robots, such as mobile helpers, physical assistants, and person carriers. It does not cover every service or humanoid robot.
- TermAdvanced
ISO 25785-1
A safety standard still in draft for robots that must actively balance to stay upright, including two-legged humanoids. It fills a gap existing standards leave open.
- TermAdvanced
ISO/TS 15066
The technical guidance for safe cobot use. It defines four modes: safety-rated monitored stop, hand guiding, speed and separation monitoring, and power and force limiting, with force limits for contact.
- TermIntermediate
Kinematics
The study of motion in terms of positions, angles, and speeds, without the forces involved. It describes how a robot's joints move to place a hand or foot.
- TermBeginner
Large language model
An AI trained on huge amounts of text to predict likely words, which lets it answer questions and write. It handles language, not physical action.
- TermIntermediate
Latency
The delay between an input and a response. In robots, high latency is dangerous because a slow reaction can mean a collision.
- TermBeginner
LiDAR
A sensor that fires laser pulses and times their return to build a precise 3D map of surroundings. It works in the dark but struggles in heavy rain or fog.
- TermIntermediate
Localization
A robot working out where it is within a space or map. Getting it wrong means a robot that gets lost or bumps into things.
- TermBeginner
Machine learning
The branch of AI where systems learn patterns from data instead of following hand-written rules. It is the foundation of nearly all modern AI.
- TermIntermediate
Motion planning
How a robot works out a full sequence of movements to get from where it is to where it needs to be without hitting anything.
- TermIntermediate
Moravec's paradox
The observation that tasks easy for humans, like walking and grasping, are hard for robots, while hard tasks like complex math are easy for machines.
- TermIntermediate
Multimodal model
An AI that handles more than one kind of input or output, such as text, images, and actions together. Robots need this to combine seeing, understanding, and moving.
- TermBeginner
Neural network
A computing structure loosely inspired by the brain, made of layers of simple units that adjust as they learn from data. The core of modern AI.
- TermAdvanced
NIST AI Risk Management Framework
A voluntary US government framework for identifying and managing AI risks, built around four functions: Govern, Map, Measure, and Manage. Widely used even though it is not law.
- TermIntermediate
Onboard compute
The computing power carried on the robot itself, used to process sensors and make decisions locally. Limited by size, weight, heat, and battery.
- TermBeginner
Open source
Software or models whose inner workings are shared publicly so anyone can inspect and build on them. It speeds up robotics and helps with transparency.
- TermIntermediate
Over-the-air update
New software sent to a robot remotely, the way phones update. It fixes bugs and adds features, but also means a robot's behavior can change after purchase.
- TermIntermediate
Parameters
The internal numbers a model adjusts during training, a rough measure of its size. Big language models have billions, while efficient robot models like ACT have far fewer.
- TermIntermediate
Path planning
Working out a route from one place to another that avoids obstacles, usually for a mobile robot. It is what keeps a warehouse robot away from shelves.
- TermBeginner
Perception
How a robot turns raw sensor data into a useful understanding of its surroundings, like recognizing people, objects, and free space. One of the hardest parts of robotics.
- TermIntermediate
Permanent magnet
A magnet that keeps its strength without power, making electric motors compact and efficient. The best ones use rare earth elements like neodymium and dysprosium.
- TermIntermediate
Physical AI
A term for AI that acts in the real world through robots and machines, rather than staying in software. It overlaps heavily with embodied AI.
- TermAdvanced
PID control
A simple, widely used way to keep a motor or joint on target by constantly correcting error based on the present, the past, and the predicted future.
- TermAdvanced
Policy (in reinforcement learning)
The strategy that maps what a robot senses to what it should do. Training aims to find a good policy, which today is usually a neural network.
- TermAdvanced
Power and force limiting
A cobot safety mode where the robot is designed so any contact with a person stays below injury limits, stopping if it meets too much resistance.
- TermIntermediate
Privacy by design
Building privacy protection into a system from the start: collecting the least data needed, securing it, and being clear about its use.
- TermAdvanced
Product liability
The law that holds makers and sellers responsible for harm caused by defective products. It is a main route to accountability when a robot injures someone.
- TermBeginner
Prompt
The instruction or question you give an AI to get a response. For a robot, it might be a spoken command that a model turns into actions.
- TermBeginner
Quadruped robot
A robot that walks on four legs, like a dog, which gives it stability on rough ground. Used for inspection, security, and exploring hazardous sites.
- TermIntermediate
Rare earth elements
A group of 17 metals vital for the strong magnets in robot and electric vehicle motors. China dominates their mining and refining, and its recent export controls have strained supply.
- TermAdvanced
Real-time computing
Computing that guarantees a response within a strict time limit, so a robot's control loop never lags dangerously. Missing a deadline counts as a failure.
- TermIntermediate
Red teaming
Deliberately trying to make an AI or robot fail or misbehave, so weaknesses are found before real users hit them. Borrowed from security testing.
- TermIntermediate
Reinforcement learning
Teaching a system by trial and error, rewarding good outcomes until it finds a strong strategy. Robots often do this in simulation, where millions of attempts are cheap.
- TermAdvanced
Reward function
The rule that tells a reinforcement learning system what counts as success. A badly designed one produces clever but useless or unsafe behavior.
- TermIntermediate
Risk assessment
A structured review of what could go wrong with a robot and how to prevent it, done before use. It is what makes a robot safe for a specific job.
- TermAdvanced
RLHF (reinforcement learning from human feedback)
Improving an AI by having people rate its outputs and using those ratings to guide further training. It is a big reason chatbots feel helpful and polite.
- TermBeginner
Robot ethics / roboethics
The study of how robots should be designed and used so they respect human safety, dignity, fairness, and rights. It guides both design and policy.
- TermIntermediate
ROS (Robot Operating System)
A widely used open-source toolkit for building robots, with ready-made parts for sensing, control, and communication. Despite the name, it is a framework, not a full operating system.
- TermAdvanced
Safety-rated monitored stop
A cobot mode where the robot safely halts when a person enters its space and resumes when they leave, without fully powering down.
- TermBeginner
Sensor
A device that lets a robot detect something about the world or itself, such as light, distance, force, or motion. Sensors are the robot's senses.
- TermIntermediate
Servo motor
A motor paired with a sensor and controller so it can move to and hold a precise position. Common in robot joints and steering.
- TermAdvanced
Sim-to-real gap
The difference between how a robot behaves in simulation and in reality. Real sensors are noisier and real physics messier, so skills learned virtually can fail in the real world.
- TermBeginner
Simulation
A virtual environment where robots are tested and trained safely and cheaply before acting in the real world. It allows huge numbers of trials without risk.
- TermAdvanced
SLAM (simultaneous localization and mapping)
How a robot builds a map of an unknown space while also tracking where it is within that map. Core to vacuum robots, drones, and self-driving cars.
- TermIntermediate
Soft robotics
Building robots from flexible materials like silicone instead of rigid metal, so they bend and squeeze. Safer around people and better with delicate objects.
- TermAdvanced
Speed and separation monitoring
A cobot mode where the robot tracks how close people are and slows or stops as they approach, keeping a safe gap while it works.
- TermBeginner
Supply chain
The full path of materials and parts needed to build a product, from raw ore to finished robot. Disruptions, like rare earth export controls, can stall production.
- TermIntermediate
Swarm robotics
Many simple robots working together toward a goal, inspired by ants and bees. The group is robust because losing one robot does not stop the job.
- TermIntermediate
Tactile sensor
A sensor that gives a robot a sense of touch, detecting contact, pressure, or texture on fingertips or grippers. It helps grasp delicate objects gently.
- TermIntermediate
Teleoperation
A person controlling a robot remotely in real time. Used for tasks too hard for autonomy or too dangerous for people, and for collecting training demonstrations.
- TermAdvanced
Tendon-driven
A design that moves joints with cables that pull like tendons, so heavy motors can sit farther away. Common in robot hands, at the cost of complexity.
- TermIntermediate
Tokens
The small chunks of text, often word pieces, that a language model reads and generates. Models are often priced by tokens, and their memory is measured in them.
- TermIntermediate
Training vs inference
Training is when a model learns from data, which is slow and expensive. Inference is when the finished model is used to answer or act, which must be fast.
- TermAdvanced
Transformer
The neural network design behind most modern AI, including chatbots and many robot models. It handles sequences by paying attention to the most relevant parts of the input.
- TermIntermediate
Trust calibration
Helping people trust a robot the right amount, matching how reliable it really is. Too little wastes a good tool, and too much leads to accidents.
- TermAdvanced
VLA (Vision-Language-Action model)
A robot AI that takes camera images and a language instruction and directly outputs the robot's actions. Examples include RT-2, pi0, Gemini Robotics, NVIDIA GR00T, and Figure's Helix.
- TermAdvanced
WAM (World Action Model)
A robot AI that predicts how the world will change and the actions to cause it. A VLA outputs actions without forecasting, and a world model forecasts without acting. Still research stage.
- TermAdvanced
World model
An AI that learns how a scene changes over time, acting like an internal simulator of reality. Examples include NVIDIA Cosmos and Google's Genie 3. Robots use it to plan before acting.
About this glossary
Sources and a note on time-sensitive facts
Chatbot rankings, market share, rare earth export controls, data center energy figures, draft standards, and the list of leading robot models change quickly. They are accurate as of September 2026 and are reviewed regularly.
- Momentic, chatbot market share using Similarweb data (May 2026)
- Gartner press release on data center electricity consumption (June 10, 2026)
- International Energy Agency, Rare Earth Elements analysis; European Parliament EPRS briefing on China's export restrictions
- S&P Global, rare earth supply bottlenecks in 2026
- Wikipedia and arXiv surveys on Vision-Language-Action models; Physical Intelligence (pi0), NVIDIA GR00T, and Figure Helix
- Zhao et al., ALOHA and ACT paper; Hugging Face LeRobot ACT documentation
- NVIDIA glossary on World Action Models; LaWAM and ABot-M0.5 arXiv papers
- ISO.org on ISO 25785-1; Automate.org and Engineering.com on ISO 10218 and ISO/TS 15066
- European Commission and Council of the EU on the AI Act and Machinery Regulation
- NIST AI Risk Management Framework 1.0 (January 2023)
- Xinhua and Robotics & Automation News on China's humanoid robot standard system (February 2026)
- SAE International J3016 levels of driving automation
Get in touch.
Connect with our team to explore collaboration opportunities or learn more about our initiatives.