Robot
Last reviewed
Sources
50 citations
Review status
Source-backed
Revision
v4 · 6,923 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
50 citations
Review status
Source-backed
Revision
v4 · 6,923 words
Add missing citations, update stale details, or suggest a clearer explanation.
A robot is a programmable machine that senses its environment, computes a decision, and acts on the physical world through a closed loop of perception, planning, and actuation. The international robotics vocabulary standard ISO 8373:2021 defines a robot as "a programmed actuated mechanism with a degree of autonomy, to perform locomotion, manipulation or positioning."[49] Robots can be guided by an external control device or have a control system embedded within them. They typically combine sensing, computation, and actuation in a single physical agent that can perceive its environment, decide what to do, and then act on the world.
Robots range from large fixed manipulators that weld car bodies in factories to centimeter-scale flying drones, soft silicone grippers, autonomous undersea vehicles, surgical assistants such as the da Vinci system, and bipedal humanoid platforms designed to work in spaces built for people. The defining property is not appearance but the closed loop between perception, decision, and physical action. A pure software agent is not a robot. A washing machine that simply runs a fixed timer is generally not considered a robot either, because it has no real perception of what is happening inside the drum.
The word robot is sometimes confused with the field of robotics. A robot is the physical machine. Robotics is the engineering discipline that designs, builds, programs, and studies robots, drawing on mechanical engineering, electrical engineering, computer science, control theory, and increasingly modern artificial intelligence. Since around 2022 the boundary between robotics and AI has narrowed considerably, with large neural networks trained on internet-scale data being adapted to control physical robots through vision-language-action models and the broader research program of embodied AI.
The word "robot" was introduced to the world by the Czech writer Karel Čapek in his 1920 stage play R.U.R., short for Rossumovi Univerzální Roboti (Rossum's Universal Robots). The play was published in 1920 and premiered at the National Theatre in Prague on 25 January 1921.[3] In R.U.R. the "robots" are mass-produced biological workers grown in vats rather than the metallic machines that later became iconic, but the dramatic core, artificial workers built to serve humans who eventually rebel, fixed the cultural template for almost a century of science fiction.
In a 1933 article in the Czech newspaper Lidové noviny, and in correspondence with the Oxford English Dictionary, Karel Čapek explained that he did not actually invent the word himself.[4] He had wanted to call the artificial workers laboři (from the Latin labor), but he found the word too bookish. He asked his brother, the painter and writer Josef Čapek, who suggested roboti.[32] The Czech word robota means forced labor, drudgery, or the kind of compulsory peasant work that survived in parts of central Europe into the nineteenth century. Through translations of R.U.R. into more than thirty languages within a few years of its premiere, robot entered English and most other European languages with essentially the same meaning.[4]
The idea of an artificial servant or self-moving statue is much older than the word robot. Mechanical automata, devices that perform a fixed sequence of motions through clever mechanisms but without sensing or programmability, are documented across many ancient cultures.
In classical antiquity, Greek and Hellenistic engineers built water clocks, hydraulic theaters, and trick vessels. Heron of Alexandria, in the first century, described programmable carts driven by falling weights and string-controlled cams, mechanical singing birds, and an early steam-driven device known as the aeolipile. Medieval Islamic engineers extended this tradition. The thirteenth-century inventor Al-Jazari, in his Book of Knowledge of Ingenious Mechanical Devices, described a wide range of automata, including a programmable musical band on a boat, hand-washing machines, and elaborate water clocks that used cams, valves, and segmental gears.
In 1495 the Italian polymath Leonardo da Vinci sketched designs for a mechanical knight, an armored humanoid driven by pulleys and cables that could sit, stand, and move its arms. Around 1737 the French inventor Jacques de Vaucanson exhibited a series of automata in Paris, including a flute player and the famous Digesting Duck (1739), which appeared to eat grain and excrete waste, although the digestion was simulated.
In 1898 the Serbian-American inventor Nikola Tesla demonstrated a radio-controlled boat in Madison Square Garden, calling the technique teleautomatics. His U.S. patent 613,809 of November 1898, titled "Method of and Apparatus for Controlling Mechanism of Moving Vessels or Vehicles," is regarded as the foundational patent for radio remote control and a conceptual ancestor of the modern teleoperated robot and unmanned aerial vehicle.
In 1920 Čapek's R.U.R. introduced the word robot. In March 1942 the American science fiction author Isaac Asimov, writing in Astounding Science Fiction, published the short story "Runaround," which contains the first explicit statement of his Three Laws of Robotics.[6] Asimov later credited a December 1940 conversation with the editor John W. Campbell, Jr. for sharpening the formulation.[5] Although the Three Laws are fictional, they shaped public expectations about how autonomous machines should be governed and remain a touchstone in modern robot ethics debates.
In 1948 and 1949 the British neurophysiologist W. Grey Walter built two small three-wheeled robots named Elsie and Elmer, sometimes called "tortoises," that used analog electronics, a photocell, and a touch sensor to seek out light sources and avoid obstacles. They are often cited as the first electronic autonomous robots in the modern sense.
The first programmable industrial robot was the Unimate, designed by the American inventor George Devol. Devol filed a patent application titled "Programmed Article Transfer" on 10 December 1954; U.S. patent 2,988,237 was granted on 13 June 1961.[7] In 1956 Devol partnered with the engineer and entrepreneur Joseph Engelberger, and together they founded Unimation, the first robotics company, in 1956. The first Unimate arm was installed on a General Motors die-casting line at the Inland Fisher Guide plant in Ewing Township, New Jersey (often described as Trenton), in 1961, where it lifted hot metal castings from the machine and stacked them, a job that was hot, monotonous, and dangerous for human workers.[8]
The 1960s and 1970s saw a series of important firsts. From 1966 to 1972, researchers at the Stanford Research Institute (now SRI International) built Shakey, the first general-purpose mobile robot able to reason about its own actions.[9] Funded by DARPA and led by Charles Rosen, Nils Nilsson, and Peter Hart, Shakey combined a TV camera, a triangulating range finder, bump sensors, a radio link, and a remote computer running planning software. The project produced the A* search algorithm, the visibility graph method, and the STRIPS automated planning representation, all foundational to artificial intelligence and modern robotics.[10] In 1973 the German firm KUKA introduced FAMULUS, the world's first six-axis industrial robot driven by electric motors rather than hydraulic actuators, a design that became the dominant kinematic configuration of articulated industrial robots.[11]
In 1986 Honda began a secret humanoid research program with a robot called E0. The program led to a series of "P" prototypes, culminating in the Honda P2 of December 1996, the first self-contained, autonomous, fully bipedal humanoid robot.[12] P2 stood about 182 centimeters tall, weighed roughly 210 kilograms, and demonstrated independent walking, stair climbing, and pushing a cart. In November 2000 Honda revealed ASIMO, a much smaller and lighter humanoid (originally about 120 centimeters and 43 kilograms) that became the public face of humanoid robotics for more than a decade.[13]
In May 1999 Sony released the AIBO ERS-110 in Japan and the United States, a quadrupedal entertainment robot dog with adaptive learning and personality features.[14] The first batch of 5,000 units sold out online in twenty minutes. In September 2002 the American company iRobot launched the Roomba, a disc-shaped robotic vacuum cleaner that became the first mass-market home robot, with one million units sold by 2004.[15]
The DARPA Grand Challenge for autonomous ground vehicles, first held in 2004, accelerated work on self-driving systems and produced the team that became Google's self-driving program. In 2013 Boston Dynamics unveiled the first-generation Atlas, a hydraulically actuated humanoid funded by DARPA for the DARPA Robotics Challenge inspired by the 2011 Fukushima Daiichi nuclear disaster.[18] In 2016 Boston Dynamics introduced Spot, a battery-powered quadruped that became commercially available in June 2020 at a list price of around $74,500.[20]
From 2021 onward a wave of new humanoid platforms appeared, driven both by maturing electric actuator and battery technology and by rapid advances in machine learning. Tesla announced its Optimus humanoid project at AI Day on 19 August 2021, and showed an untethered prototype called "Bumble C" walking on stage on 30 September 2022.[21] The startup Figure AI was founded in 2022 and showed Figure 01 in 2023, raising $675 million in February 2024 from investors including Microsoft, OpenAI, NVIDIA, Jeff Bezos, and Intel Capital.[22] 1X Technologies (formerly Halodi Robotics) released the NEO Beta in August 2024.[23] Apptronik announced its Apollo humanoid in 2023 and entered a commercial pilot with Mercedes-Benz in March 2024.[25] Agility Robotics reached a milestone the same year by deploying its Digit humanoid in commercial fulfillment work for GXO Logistics, generally regarded as the first paid commercial deployment of a humanoid robot.[24] Unitree of China released the H1 humanoid in 2023 and the smaller G1 in 2024, both at price points well below Western competitors.
In April 2024 Boston Dynamics retired its iconic hydraulic Atlas and unveiled an entirely electric replacement built for commercial use under Hyundai Motor Group ownership.[19] The same period brought a parallel revolution in robot software. Google DeepMind released RT-1 in December 2022 and the larger RT-2 in July 2023, the first widely cited vision-language-action model trained jointly on web data and robot trajectories.[26] The startup Physical Intelligence released the π0 foundation model in October 2024 and closed a $400 million Series A on 4 November 2024 at a reported $2.4 billion valuation, with backing from Bezos, OpenAI, Thrive Capital, Lux Capital, and Bond Capital.[27][28] NVIDIA introduced Project GR00T at GTC 2024 and the GR00T N1 generalist humanoid foundation model in 2025.
The humanoid build-out accelerated sharply through 2025 and into 2026, with several startups reaching valuations in the tens of billions of dollars and the first Chinese embodied-AI companies preparing public listings. At the same time, robot software consolidated around a small number of large vision-language-action foundation models.
On 16 September 2025 Figure AI announced that it had committed more than $1 billion in its Series C round at a $39 billion post-money valuation, a roughly fifteen-fold increase over the $2.6 billion valuation it carried in early 2024.[33] The round was led by Parkway Venture Capital with participation from Brookfield, NVIDIA, Intel Capital, LG Technology Ventures, Salesforce, Qualcomm Ventures, and others.[33] In October 2025 Figure unveiled Figure 03, its third-generation humanoid redesigned for high-volume manufacturing at its BotQ facility and for home as well as commercial use, running the company's in-house Helix VLA.[33]
On 11 February 2026 Apptronik closed a $520 million Series A extension, lifting the total Series A to about $935 million and its post-money valuation to roughly $5.3 billion.[34] The extension included Google and Mercedes-Benz among its backers; Apptronik builds its Apollo humanoid for tasks such as unloading trailers and machine tending, and is a partner of Google DeepMind, GXO, and Mercedes-Benz.[34] Physical Intelligence raised about $600 million in November 2025 at a $5.6 billion valuation, in a round led by Alphabet's CapitalG with participation from Lux Capital, Thrive Capital, and Jeff Bezos.[35]
1X Technologies shifted its focus toward the consumer home market, opening pre-orders for its NEO home humanoid on 28 October 2025 at a price of $20,000 outright or a $499 per month subscription, with deliveries planned for 2026.[36] In December 2025 the company announced a partnership with the investment firm EQT to make up to 10,000 NEO robots available to EQT portfolio companies for industrial use between 2026 and 2030, broadening NEO beyond the home.[37]
In China, Unitree launched the lower-cost R1 humanoid (priced from roughly $5,900) in July 2025 and, according to the Shanghai Stock Exchange, received listing-committee approval for an initial public offering on 1 June 2026, positioning it to become one of the first publicly traded "embodied AI" companies on China's A-share market; the company had filed in March 2026 for an offering of about 4.2 billion yuan (roughly $610 million).[38] The Shanghai-based rival AgiBot (Zhiyuan Robotics) said its 10,000th robot rolled off the line by March 2026 and reported plans for a Hong Kong listing later in the year.[39] According to market trackers, Chinese manufacturers accounted for the large majority of humanoid unit shipments in 2025, with AgiBot and Unitree each shipping several thousand units, well ahead of U.S. rivals such as Figure and Tesla, which each shipped only a few hundred or fewer.[39]
Tesla's Optimus program remained in a research and data-collection phase through this period. On its January 2026 (fourth-quarter 2025) earnings call, chief executive Elon Musk said the Optimus units operating inside Tesla factories were there to generate training data rather than to perform productive work, and the company subsequently pushed the public reveal of its next-generation ("V3") design toward the middle of 2026, with a Fremont production line targeted at an eventual capacity of about one million units per year.[40]
There is no single accepted definition of "robot." The most widely used international reference is ISO 8373:2021, the robotics vocabulary standard, which defines a robot as "a programmed actuated mechanism with a degree of autonomy, to perform locomotion, manipulation or positioning," where autonomy is "the ability to perform intended tasks based on current state and sensing, without human intervention."[49] The same standard defines an industrial robot as an "automatically controlled, reprogrammable, multipurpose manipulator, programmable in three or more axes," and a service robot as one that "performs useful tasks for humans or equipment excluding industrial automation applications." The boundary between a robot and a sophisticated appliance, computer-controlled machine tool, or remote-controlled toy is not always sharp.
A common practical taxonomy classifies robots by their physical form and primary application:
| Class | Typical examples | Primary use |
|---|---|---|
| Industrial robot | KUKA KR series, FANUC R-2000, ABB IRB | Welding, painting, assembly, palletizing, machine tending in factories |
| Collaborative robot (cobot) | Universal Robots UR5/UR10, FANUC CRX, ABB GoFa | Light assembly and pick-and-place alongside humans without safety cages |
| Mobile robot (AGV/AMR) | Locus, Geek+, Symbotic, Amazon Robotics drives | Warehouse transport, intralogistics |
| Humanoid robot | ASIMO, Atlas, Optimus, Figure, Digit, NEO | Research; emerging commercial work in factories and homes |
| Quadruped | Boston Dynamics Spot, Unitree Go2, ANYmal | Inspection, security, construction monitoring, research |
| Aerial robot (UAV/drone) | DJI Matrice, Skydio X10, military fixed-wing UAVs | Photography, mapping, delivery, surveillance, agriculture |
| Underwater robot (AUV/ROV) | Saab Sabertooth, Bluefin-21, Boeing Echo Voyager | Pipeline inspection, oceanography, search and recovery |
| Surgical robot | Intuitive Surgical da Vinci, Medtronic Hugo, CMR Versius | Minimally invasive surgery |
| Service robot | Roomba, Bear Robotics Servi, Knightscope | Cleaning, food service, security |
| Soft robot | Octopus-inspired grippers, pneumatic crawlers | Delicate handling, conformable inspection |
| Swarm robot | Kilobot, Harvard TERMES | Research into distributed coordination |
| Telepresence robot | Double, OhmniLabs | Remote presence in offices, hospitals, schools |
| Educational robot | LEGO Mindstorms, VEX, Sphero | STEM teaching, competitions |
A single physical platform can fit several categories. Spot is both a quadruped and a service or inspection robot. The da Vinci is both a surgical and a teleoperated robot, since it is master-slave controlled rather than autonomous.
A modern robot is essentially a tightly integrated assembly of mechanical, electrical, sensing, computing, and software subsystems. The typical major components are summarized below.
| Subsystem | Function | Common technologies |
|---|---|---|
| Mechanical structure | Provides the body, links, joints, and chassis | Aluminum, steel, carbon fiber, machined or 3D-printed parts |
| Actuators | Convert electrical, hydraulic, or pneumatic power into motion | Brushless DC motors with planetary or harmonic-drive gearing; quasi-direct-drive motors; hydraulic cylinders; pneumatic muscles; shape-memory alloys |
| Sensors | Measure internal state and the external world | Joint encoders, force/torque sensors, IMUs, RGB and depth cameras, LiDAR, radar, sonar, microphones, tactile arrays |
| Compute | Runs perception, planning, and control software | Microcontrollers (e.g. STM32) for low-level loops; embedded GPUs such as NVIDIA Jetson Orin or Thor; x86 industrial PCs; in some cases Apple silicon or AMD modules |
| Power | Supplies electrical or fluid energy | Lithium-ion batteries, hydraulic pumps, tethered power, fuel cells (rare) |
| Communication | Connects the robot to operators and other systems | Wi-Fi, 5G, Ethernet, CAN bus, EtherCAT, real-time fieldbuses |
| End-effectors | Tools at the manipulator tip that interact with the world | Two- and three-finger grippers, multi-fingered hands, suction cups, welding torches, drills, scalpels |
| Software stack | Coordinates everything | Robot Operating System (ROS / ROS 2), proprietary middleware, motion planners, controllers, learned policies |
Classical robot software is organized as a pipeline of perception, state estimation, planning, and control. Modern AI-driven robots increasingly replace large parts of that pipeline with learned end-to-end policies trained on demonstrations or in simulation.
The classical organizing principle of robot autonomy is the sense-plan-act (SPA) paradigm, also called the deliberative or hierarchical paradigm.[50] In it, the robot proceeds top-down through three primitives in sequence: it senses the environment, plans by updating an internal world model and searching for an action, then acts by executing that plan, before repeating the cycle. Shakey, the SRI mobile robot of 1966-1972, is the canonical early example, and the sense-plan-act loop dominated robot architectures into the 1980s.[50]
In 1986 the MIT roboticist Rodney Brooks challenged this approach with his subsumption architecture, arguing that rebuilding a full world model before every action made robots too slow and brittle for the real world.[50] His reactive paradigm instead couples sensing directly to behavior with little or no central planning, while later hybrid architectures reintroduce a deliberative planning layer that selects and sequences reactive behaviors. These three paradigms, deliberative, reactive, and hybrid, remain the standard framework for describing how a robot turns perception into action, and the modern learned policies described below can be seen as a fourth, data-driven way of closing the same loop.[50]
The industrial robot market is dominated by four companies, sometimes called the "Big Four," along with several other major players. The table below lists representative manufacturers.
| Company | Headquarters | Founded | Notable products and notes |
|---|---|---|---|
| FANUC | Oshino, Yamanashi, Japan | 1972 (independent) | World's largest industrial robot maker; ROBOSHOT, R-2000, CRX cobots; characteristic yellow paint |
| ABB Robotics | Vasteras, Sweden / Zurich, Switzerland | 1988 (ABB merger; ASEA robotics from 1974) | IRB series, YuMi dual-arm cobot, GoFa |
| KUKA | Augsburg, Germany | 1898 (acetylene); robotics from 1973 | KR series, LBR iiwa cobot; acquired by China's Midea Group, deal completed January 2017[30] |
| Yaskawa Motoman | Kitakyushu, Japan | 1915 (Yaskawa); Motoman robotics from 1977 | MOTOMAN-L10 was the first all-electric Japanese industrial robot; HC series cobots |
| Universal Robots | Odense, Denmark | 2005 | Pioneered the modern collaborative robot with the UR5 (2008); now part of Teradyne[29] |
| Kawasaki Robotics | Akashi, Japan | Robotics from 1969 | Licensed Unimate technology in 1969, the first robotics maker in Japan |
| Stäubli | Pfäffikon, Switzerland | 1892 (textile); robotics from 1982 | TX series, food and pharma applications |
| Epson Robots | Suwa, Japan | Robotics from 1980 | High-speed SCARA arms |
| DENSO Robotics | Kariya, Japan | Robotics from 1967 (in-house); external from 1990s | Compact assembly robots |
| Mitsubishi Electric | Tokyo, Japan | Robotics from 1980 | MELFA series |
| Comau | Turin, Italy | 1973 | Body-in-white welding for automotive |
According to the International Federation of Robotics (IFR), the global stock of industrial robots in operation reached approximately 4.28 million units at the end of 2023 and grew further in 2024.[1] China alone accounts for roughly half of all annual installations.[2]
Outside the factory, service robots have become a large and growing segment. The most successful consumer robot to date is the iRobot Roomba robotic vacuum, launched in September 2002, which had sold tens of millions of units by the early 2020s.[15] Amazon announced an agreement to acquire iRobot in August 2022 for $1.7 billion, but the deal was abandoned in January 2024 after antitrust scrutiny in the European Union. Robotic lawn mowers such as the Husqvarna Automower (introduced in the mid-1990s and steadily improved since), as well as window-cleaning and pool-cleaning robots, occupy similar niches.
In humanoid-styled service robotics, SoftBank Robotics introduced Pepper in June 2014.[31] Pepper is a 1.2-meter wheeled humanoid designed for emotional interaction in retail, hospitality, and education settings. Production was paused in 2020 and confirmed discontinued in mid-2021 after only about 27,000 units had been built.[31] SoftBank Robotics also produced the smaller NAO educational humanoid, originally developed by the French company Aldebaran Robotics from 2008.
Sony's AIBO robot dog, launched in 1999 and discontinued in 2006, was relaunched as the ERS-1000 in Japan in January 2018 and the United States in September 2018.[14] Smaller social robots such as Anki Cozmo and Vector (Anki shut down in 2019; Vector was later revived under Digital Dream Labs) and the desktop companion Eilik from Energize Lab have continued the consumer-robot tradition.
Warehouse and logistics automation became one of the largest real-world robot deployments of the mid-2020s. In July 2025 Amazon said it had deployed its one-millionth warehouse robot, with a milestone unit installed at a fulfillment center in Japan, and that mobile robots now assist a large share of its order fulfillment.[41] Alongside the hardware milestone Amazon introduced DeepFleet, a generative-AI coordination model trained on its warehouse operations data to route mobile robots more efficiently, which the company said improved fleet travel efficiency by about 10 percent; in May 2025 it also unveiled Vulcan, a stowing robot with a force-sensing gripper that can "feel" the items it handles.[41]
The period from roughly 2021 to the present has seen an unprecedented surge of investment and progress in general-purpose humanoid robots, driven by the convergence of cheaper electric actuators, higher-density batteries, and powerful machine-learning models. The table below summarizes the most active programs.
| Company | Robot | First public reveal | Notes |
|---|---|---|---|
| Tesla | Optimus | Concept Aug 2021; "Bumble C" prototype Sept 30, 2022 | Iterations announced as Gen 1, Gen 2, and Gen 3; intended for both factory and consumer use; next-gen reveal pushed to mid-2026[40] |
| Figure AI | Figure 01 / 02 / 03 | 01 unveiled Oct 2023 | Raised $675M in Feb 2024; collaboration with OpenAI announced 2024 (modified later); 02 unveiled Aug 2024; 03 unveiled Oct 2025; Series C >$1B at $39B in Sept 2025[33] |
| 1X Technologies | EVE, NEO | NEO Beta Aug 2024 | Backed by OpenAI Startup Fund; NEO home robot pre-orders opened Oct 2025 at $20,000; targeting home use; based in Norway and California[36] |
| Apptronik | Apollo | Aug 2023 | Mercedes-Benz pilot announced March 2024; collaboration with NVIDIA on Project GR00T; ~$935M Series A at ~$5.3B by Feb 2026[34] |
| Agility Robotics | Digit | Cassie (predecessor) 2017; Digit v1 2019; latest revisions 2024 | First humanoid in commercial paid deployment, with GXO at a Spanx warehouse in Georgia (2024) |
| Boston Dynamics | Atlas (electric) | Hydraulic Atlas 2013, retired April 16, 2024; Electric Atlas April 17, 2024 | Owned by Hyundai Motor Group; first humanoid product effort; Large Behavior Model demo with Toyota Research Institute Aug 2025[42] |
| Sanctuary AI | Phoenix | May 2023; 8th-generation Jan 2025 | Vancouver-based; partnership with Magna International; pioneering hydraulic dexterous hands |
| Unitree | H1, H2, G1, R1 | H1 mid-2023; G1 May 2024; R1 July 2025 | Aggressive pricing, with G1 starting around $16,000 and R1 from about $5,900; Shanghai IPO approved June 2026[38] |
| Xiaomi | CyberOne | Aug 2022 | Demonstration platform; not a commercial product |
| Fourier Intelligence | GR-1 | Sept 2023 | China-based; targeted at research and rehabilitation |
| UBTech | Walker, Walker S | Walker 2018; Walker S in 2024 | Pilots with Chinese automakers including BYD and Geely |
| XPeng (Xpeng) | Iron | Nov 2024 | Demonstrated walking the catwalk at AI Day 2024 |
| AgiBot (Zhiyuan) | Yuanzheng, Lingxi | 2023 | Shanghai-based; 10,000th robot built by March 2026; Hong Kong IPO planned[39] |
The field is fast-moving and not all programs will reach commercial viability. Several teams have already reorganized, paused, or pivoted; for example, Sanctuary AI publicly emphasized data-collection use cases for Phoenix in 2025, and various Chinese players have consolidated under municipal or industry-driven humanoid robotics innovation centers.
The integration of artificial intelligence with robots is the most active area of robotics research and development today, and is one of the central reasons for renewed investor interest in the field.
Classical robotics relies on hand-engineered models. Forward and inverse kinematics describe the relationship between joint angles and end-effector pose. Dynamics use Newton-Euler or Lagrangian formulations to model torques and accelerations. Trajectory planning uses sampling-based algorithms such as Probabilistic Roadmaps (PRM) and Rapidly-exploring Random Trees (RRT and RRT*). Localization and mapping in mobile robots use techniques such as Extended Kalman Filtering, particle filters, and graph-based simultaneous localization and mapping (SLAM). Low-level control is dominated by PID, computed-torque, and operational-space controllers. These methods remain essential and are still the foundation of most safety-critical industrial systems.
Since roughly 2015, machine learning, particularly deep reinforcement learning and imitation learning, has begun to replace or augment classical pipelines for tasks where it is hard to write a model by hand, such as dexterous manipulation, locomotion over rough terrain, and visual recognition. Reinforcement-learning algorithms such as Deep Deterministic Policy Gradient (DDPG), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC) train policies in simulation and transfer them to physical robots through techniques such as domain randomization. Imitation learning trains policies from human demonstrations, often collected via teleoperation or virtual-reality rigs. Algorithms such as Behavioral Cloning, Diffusion Policy, and Action Chunking Transformer (ACT) are widely used.
The most striking recent development is the rise of robot foundation models, large neural networks pretrained on broad data that can then be specialized for specific robots and tasks. As Physical Intelligence put it when releasing π0 in October 2024, "In the same way that LLMs provide a foundation model for language, these generalist robot policies will provide a robot foundation model for physical intelligence."[27]
| Model | Organization | Year | Description |
|---|---|---|---|
| RT-1 | Google Robotics | December 2022 | Transformer policy trained on 130,000 episodes from 13 robots |
| RT-2 | Google DeepMind | July 2023 | First widely cited vision-language-action (VLA) model; built on PaLI-X / PaLM-E[26] |
| Open X-Embodiment / RT-X | Cross-institution consortium | October 2023 | Open dataset of 1M+ trajectories from 22 robot embodiments |
| OpenVLA | Stanford / collaborators | June 2024 | Open-source 7B-parameter VLA based on Llama 2 |
| π0 (Pi-Zero) | Physical Intelligence | October 2024 | Generalist VLA trained on data from 8 distinct robots, controlling 7+ robot types over dozens of tasks[27] |
| GR00T N1 | NVIDIA | March 2025 | Open foundation model targeted at humanoid robots[43] |
| Helix | Figure AI | February 2025 | In-house VLA powering Figure humanoids |
| Gemini Robotics | Google DeepMind | March 2025 | VLA built on Gemini 2.0, paired with the Gemini Robotics-ER embodied-reasoning model[44] |
| π0.5 (Pi-0.5) | Physical Intelligence | April 2025 | Successor to π0 with open-world generalization; cleans kitchens and bedrooms in unseen homes[45] |
| Gemini Robotics On-Device | Google DeepMind | June 2025 | Local VLA that runs on the robot; first DeepMind robot model offered for fine-tuning[46] |
| Gemini Robotics 1.5 / ER 1.5 | Google DeepMind | September 2025 | Agentic ER+VLA stack that plans multi-step tasks and transfers skills across embodiments[47] |
The April 2025 π0.5 model from Physical Intelligence, trained with a technique the authors call knowledge insulation and co-training on heterogeneous data (multiple robots, web data, semantic subtask prediction, and low-level actions), demonstrated a mobile manipulator cleaning kitchens and bedrooms in homes that were never seen during training, a benchmark for open-world generalization.[45] NVIDIA followed GR00T N1 with the updated, commercially licensable GR00T N1.5 and later checkpoints (N1.6, N1.7) in 2025, distributed for deployment on its Jetson Thor (Blackwell-based) robot computer announced the same year.[43] Google DeepMind's Gemini Robotics line progressed from the March 2025 cloud model to an on-device version in June 2025 and to the agentic Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 in September 2025, the latter able to reason about the physical world, call digital tools, and create multi-step plans, with the VLA able to transfer skills across different robot bodies including Apptronik's Apollo.[44][46][47]
A related development is the maturation of high-fidelity simulators such as MuJoCo (originally developed by Emo Todorov, open-sourced by DeepMind in 2021), NVIDIA Isaac Sim and Isaac Lab, the Gazebo and Gazebo Sim ecosystem associated with ROS, and the Genesis simulator released in 2024. These tools enable large-scale parallel reinforcement learning and synthetic data generation for sim-to-real transfer.
Dexterous manipulation, the ability to grasp and reorient arbitrary objects, has long been a defining challenge in robotics. Modern learned policies, often trained from a combination of human demonstrations and reinforcement learning in simulation, have produced impressive results on tasks such as in-hand cube reorientation, cloth folding, and assembly. Locomotion has likewise been transformed: quadrupeds such as Spot, ANYmal, and the Unitree platforms now traverse stairs, mud, and snow using neural network controllers trained almost entirely in simulation. In August 2025 Boston Dynamics and the Toyota Research Institute demonstrated a single Large Behavior Model controlling the electric Atlas through a long sequence of combined whole-body manipulation and locomotion tasks, an approach the partners said let them add new behaviors without hand-coding each one.[42]
The Three Laws of Robotics formulated by Isaac Asimov in 1942 are works of fiction, but they have powerfully shaped public expectations and academic discussion about how autonomous machines should be governed.[5] The actual safety of robots in the real world is governed by engineering standards, occupational-safety regulations, and emerging product-liability frameworks.
The main international standard for industrial robot safety is ISO 10218, parts 1 and 2, which covers the design and integration of industrial robots. Collaborative robots, which work in shared spaces with humans, are additionally governed by ISO/TS 15066, published in 2016, which specifies allowable contact forces and pressures for different parts of the human body. In Europe, the EU Machinery Regulation (Regulation (EU) 2023/1230, replacing the earlier Machinery Directive 2006/42/EC) sets essential health and safety requirements for placing machines, including robots, on the EU market. In the United States, OSHA, ANSI/RIA R15.06, and (for medical devices) the FDA play parallel roles.
For autonomous robots and vehicles, evolving liability frameworks ask who is responsible when an autonomous system causes harm: the operator, the manufacturer, or the supplier of a learned policy. The European Union's AI Act, adopted in 2024, classifies certain robotic systems as high-risk AI and imposes documentation and conformity-assessment requirements. The use of fully autonomous weapons, often called Lethal Autonomous Weapons Systems (LAWS) or popularly "killer robots," has been the subject of ongoing United Nations Convention on Certain Conventional Weapons (CCW) discussions in Geneva since 2014, with a coalition of states and non-governmental organizations such as the Campaign to Stop Killer Robots calling for a binding treaty.
The global robotics industry is now firmly established as one of the largest segments of advanced manufacturing equipment. According to the International Federation of Robotics (IFR) World Robotics 2024 report, the operational stock of industrial robots reached approximately 4,281,585 units worldwide at the end of 2023, a roughly 10 percent increase year-on-year.[1] The 2025 edition of the report puts the operational stock at approximately 4.66 million units (4,664,000) at the end of 2024, a 9 percent year-on-year increase, with about 542,000 new industrial robots installed during 2024, the second-highest annual total on record.[2] Asia accounted for about 74 percent of new installations, with China by far the single largest national market at roughly 54 percent of global installations.[2] Announcing the figures, IFR President Takayuki Ito said: "The transition of many industries into the digital and automated age has been marked by a huge surge in demand."[2] Other major adopters include Japan, the United States, South Korea, and Germany. Robot density (installed robots per 10,000 manufacturing employees) is highest in South Korea, Singapore, China, Germany, and Japan. The IFR projected that global installations would rise by about 6 percent to roughly 575,000 units in 2025, and reported that China's operational stock passed the two-million mark in 2024, the largest of any country.[2]
The service-robot market is smaller in unit revenue but growing rapidly, driven by warehouse automation (Amazon Robotics, Symbotic, AutoStore, Geek+, and others), commercial cleaning, food service, healthcare, and consumer vacuums. Investment in humanoid-robot startups since 2023 has been historic. Notable rounds include Figure AI's $675 million Series B in February 2024 at a reported $2.6 billion post-money valuation,[22] 1X Technologies' $100 million Series B in January 2024,[23] Physical Intelligence's $400 million Series A in November 2024 at roughly $2.4 billion,[28] Apptronik's $350 million Series A in February 2025, and continued large rounds for Skild AI, Covariant, and several Chinese players supported by municipal innovation funds. By late 2025 and early 2026 the largest rounds had grown by an order of magnitude: Figure AI's Series C exceeded $1 billion at a $39 billion valuation in September 2025,[33] Physical Intelligence raised about $600 million at a $5.6 billion valuation in November 2025,[35] and Apptronik's Series A reached about $935 million at roughly $5.3 billion by February 2026.[34] Bank forecasts framed the long-term opportunity in trillions of dollars; Morgan Stanley, for example, projected in 2025 that the humanoid market and its surrounding ecosystem could approach $5 trillion by 2050.[48]
| Year | Name | Origin | Significance |
|---|---|---|---|
| 1738-1739 | Vaucanson's Digesting Duck | Jacques de Vaucanson, France | Iconic mechanical automaton |
| 1898 | Tesla's teleautomaton | Nikola Tesla, USA | First demonstrated radio-controlled vehicle |
| 1948-1949 | Elsie and Elmer ("tortoises") | W. Grey Walter, UK | Early autonomous electronic robots |
| 1961 | Unimate | George Devol / Unimation, USA | First programmable industrial robot deployed at GM |
| 1966-1972 | Shakey | SRI International, USA | First mobile robot with reasoning ability |
| 1973 | KUKA FAMULUS | KUKA, Germany | First six-axis electrically driven industrial robot |
| 1996 | Honda P2 | Honda, Japan | First fully self-contained autonomous bipedal humanoid |
| 1999 | Sony AIBO ERS-110 | Sony, Japan | First mass-produced consumer entertainment robot |
| 2000 | Honda ASIMO | Honda, Japan | Public face of humanoid robotics for over a decade |
| 2000 | Intuitive Surgical da Vinci | Intuitive Surgical, USA | First FDA-cleared general-surgery robotic system |
| 2002 | iRobot Roomba | iRobot, USA | First mass-market home robot |
| 2008 | Universal Robots UR5 | Universal Robots, Denmark | First commercially successful collaborative robot |
| 2013 | Boston Dynamics Atlas (hydraulic) | Boston Dynamics / DARPA, USA | Iconic humanoid platform of the 2010s |
| 2016 | Boston Dynamics Spot | Boston Dynamics, USA | First widely deployed commercial quadruped |
| 2022 | Tesla Optimus prototype "Bumble C" | Tesla, USA | Catalyst for the modern humanoid race |
| 2024 | Agility Digit at GXO | Agility Robotics / GXO, USA | First humanoid robot in paid commercial deployment |
| 2024 | Boston Dynamics Electric Atlas | Boston Dynamics / Hyundai, USA | First Boston Dynamics humanoid product |
| 2024 | Physical Intelligence π0 | Physical Intelligence, USA | Influential generalist robot foundation model |
| 2025 | Amazon one-millionth warehouse robot | Amazon, USA / Japan | Largest deployed mobile-robot fleet, paired with the DeepFleet AI model |
Robots have long been one of the most important narrative devices in science fiction, often used to explore questions of labor, consciousness, ethics, and identity. The astromech droid R2-D2 and the protocol droid C-3PO of Star Wars (from 1977), the murderous spaceship intelligence HAL 9000 in 2001: A Space Odyssey (1968), the android Data in Star Trek: The Next Generation (from 1987), the lonely waste-compactor robot Wall-E in the 2008 Pixar film of the same name, the cynical Bender in Futurama (from 1999), and the relentless cyborg of The Terminator (1984) are among the most recognizable. These cultural touchstones have influenced public expectations of real robots, often setting them up to disappoint, since most working robots remain rigid, narrow-purpose, and clumsy compared to their fictional counterparts. The relationship runs the other way as well: many roboticists trace their interest in the field to childhood encounters with such characters.