Updated September 2026

Timeline of Robotics.

40 checkpoints, from two vacuum tubes in a plastic shell to policies trained across twenty-two different bodies. Scroll sideways.

Before the Machines

18981/40

1898

Tesla’s radio-controlled boat

Nikola Tesla demonstrates wireless remote control and calls it a borrowed mind.

At Madison Square Garden Tesla steered a small boat around a tank by radio, to an audience that largely assumed trickery or a trained animal inside. His patent describes the vessel as having a borrowed mind, and he predicted machines that would act on their own. It is the first demonstration of remote machine control and the ancestor of every teleoperated system.

Tesla, Method of and apparatus for controlling mechanism of moving vessels

Before the Machines

19212/40

1921

The word robot

Karel Capek’s play R.U.R. names the thing, and the name carries a warning.

Rossum’s Universal Robots gave us the word, from the Czech robota, meaning forced labour. Capek’s robots are manufactured workers who eventually destroy humanity, so the term arrived pre-loaded with anxiety about replacement. His brother Josef actually coined it. The play is rarely staged and the word has never been replaced.

Capek, R.U.R.

Before the Machines

19423/40

1942

Asimov writes the Three Laws

A science fiction device gives the field a vocabulary it is still arguing with.

Asimov introduced the Laws in the short story Runaround, mostly as a plot generator: every story works by finding an edge case where the rules conflict. They are not engineering specifications and were never meant to be. But they established robot safety as something to be specified rather than assumed, and the word robotics itself is his coinage.

Asimov, Runaround

Before the Machines

19494/40

1948–1949

Grey Walter’s tortoises

Two analogue robots with two neurons each produce startlingly lifelike behaviour.

Elmer and Elsie had a light sensor, a touch sensor, two vacuum tubes and two motors. With no program and no memory they approached moderate light, avoided bright light, backed away from obstacles and returned to a charging hutch when their batteries ran low. Walter’s point was that complex behaviour need not imply complex internals, an argument the field would have to relearn.

Walter, An Imitation of Life

Industrial Arms

19615/40

1961

Unimate goes to work

The first industrial robot starts unloading a die-casting machine at General Motors.

Devol patented the programmable transfer device and Engelberger turned it into a company. The Unimate weighed about two tonnes, stored its program on a magnetic drum, and handled hot die castings that regularly injured the humans who had done the job. It was hydraulic, blind and utterly repeatable, which for that task was exactly the right set of properties.

Devol, Programmed Article Transfer (US patent 2,988,237)

Industrial Arms

19696/40

1969

The Stanford Arm

Victor Scheinman designs an all-electric arm meant for computer control rather than playback.

Where the Unimate replayed recorded motions, the Stanford Arm was built to be solved. Six degrees of freedom, electric drives and a geometry chosen so the kinematics had a closed-form solution meant a computer could compute joint angles for a desired hand position. It is the direct ancestor of essentially every industrial manipulator since, including the PUMA.

Scheinman, Design of a computer controlled manipulator

Industrial Arms

19697/40

1966–1972

Shakey reasons about its world

SRI builds the first robot that plans its own actions rather than executing a script.

Shakey perceived a blocks world with a television camera, built a symbolic model of it, and used the STRIPS planner to work out action sequences achieving a goal. It moved slowly enough that observers went for coffee. Its legacy is enormous: STRIPS planning, the A* search algorithm and the Hough transform all came out of the project.

Nilsson, Shakey the Robot

Industrial Arms

19738/40

1973

WABOT-1, the first humanoid

Waseda University builds a full-scale anthropomorphic robot that walks, sees and converses.

WABOT-1 had a limb control system, a vision system and a conversation system. It could walk, grip objects, measure distances and hold a simple spoken exchange in Japanese. Its mental capability was reckoned at that of an eighteen-month-old child and each step took forty-five seconds, but it established the integrated humanoid as a research programme.

Waseda University Humanoid Robotics Institute

Industrial Arms

19789/40

1978

PUMA and the assembly line

Unimation and GM produce a programmable arm scaled to human tasks, and it becomes the standard.

The Programmable Universal Machine for Assembly was designed around a person’s reach and payload rather than around heavy foundry work. It was accurate, affordable by the standards of the day and widely cloned, and it dominated both factories and university labs for two decades. A PUMA also performed the first robot-assisted surgery in 1985.

Unimation PUMA

Industrial Arms

198110/40

1981

SCARA

A deliberately constrained arm geometry turns out to be far better at the job that matters.

Hiroshi Makino reasoned that most assembly is vertical insertion, so an arm should be compliant horizontally and stiff vertically. The Selective Compliance Assembly Robot Arm has joints that rotate only about vertical axes plus one vertical slide. Fewer degrees of freedom made it faster, cheaper and more precise for its task, and it still dominates electronics assembly.

Makino, Development of the SCARA

Behaviour and Bodies

198611/40

1986

Subsumption architecture

Rodney Brooks argues that robots should react to the world, not model it.

Classical robotics sensed, built a world model, planned and acted, and choked on the modelling step. Brooks proposed layered behaviours wired directly from sensors to actuators, with higher layers suppressing lower ones. His robots walked over rough ground with no map at all. The slogan was that the world is its own best model, and the argument split the field.

Brooks, A robust layered control system for a mobile robot

Behaviour and Bodies

198912/40

1989

Fast, cheap and out of control

Brooks proposes swarms of small simple robots instead of one large complex one.

The paper argued that planetary exploration would be better served by many disposable insect-scale robots than by a single expensive rover whose failure ends the mission. NASA did not take the bet at the time, but the reasoning shaped behaviour-based robotics, swarm research, and eventually a company: Brooks went on to found iRobot.

Brooks and Flynn, Fast, Cheap and Out of Control

Behaviour and Bodies

199013/40

1990

Passive dynamic walking

Tad McGeer builds a walker with no motors, no sensors and no control at all.

Placed on a shallow slope, McGeer’s device walked with a natural human-like gait powered only by gravity. It demonstrated that much of the dynamics of walking is in the mechanics of the limbs rather than in the controller. Designing the body well removes work from the software, which became a foundational idea in legged robotics.

McGeer, Passive dynamic walking

Behaviour and Bodies

199714/40

1997

Sojourner reaches Mars

A microwave-sized rover proves autonomous operation across an eleven-minute light delay.

Teleoperation is impossible when commands take minutes to arrive, so Sojourner had to detect hazards and choose its own local paths. It travelled about a hundred metres in total over 83 sols. The autonomy requirement it established has driven every planetary rover since, up to Perseverance choosing its own routes at speed.

NASA Mars Pathfinder

Behaviour and Bodies

200015/40

2000

ASIMO walks

Honda shows a humanoid that walks smoothly, turns while moving and climbs stairs.

Honda had worked on bipedal locomotion in secret since 1986. ASIMO used zero moment point control, keeping the ground reaction force inside the support polygon, which produced the characteristic slightly crouched gait. It was carefully choreographed and fragile off script, but it made walking humanoids look like engineering rather than speculation.

Honda ASIMO

Robots in the House

200016/40

2000

The da Vinci surgical system

A teleoperated robot gets FDA clearance and scales a surgeon’s hands down past human tremor.

The surgeon works at a console, and the system filters tremor and scales motion so a centimetre of hand movement becomes a millimetre at the instrument. Combined with wristed instruments inside small incisions, it made laparoscopic procedures feasible that hand tools could not manage. Millions of procedures later it remains the most commercially successful medical robot by a wide margin.

Intuitive Surgical da Vinci

Robots in the House

200217/40

2002

Roomba

iRobot ships a domestic robot that succeeds by being cheap and doing one thing.

The first Roomba had no map and no localisation. It bounced semi-randomly, followed walls and spiralled on dirt, using coverage-over-time instead of planning. That was enough for the task and kept the price low enough to sell in volume. It remains the only genuinely mass-market home robot, and the lesson was that constraint beats generality in consumer products.

iRobot Roomba

Robots in the House

200218/40

2002

Probabilistic robotics and SLAM

Treating localisation and mapping as one joint probability problem finally makes both tractable.

A robot needs a map to know where it is, and needs to know where it is to build a map. Simultaneous localisation and mapping solves both at once by carrying uncertainty explicitly through particle filters and factor graphs rather than pretending sensors are exact. Thrun, Burgard and Fox’s formulation underlies every autonomous vehicle, drone and cleaning robot that builds a map.

Thrun, Burgard and Fox, Probabilistic Robotics

Robots in the House

200519/40

2005

Stanley finishes the Grand Challenge

A year after nobody completed the course, five vehicles drive 132 desert miles autonomously.

In 2004 the best entry managed seven miles. In 2005 Stanford’s Stanley finished in under seven hours, using lidar, machine learning over terrain classification and probabilistic state estimation rather than hand-written rules. The DARPA challenges created the talent pool and the technical playbook that the entire self-driving industry was built from.

Thrun et al., Stanley: The robot that won the DARPA Grand Challenge

Robots in the House

200520/40

2005

BigDog stays upright

Boston Dynamics builds a quadruped that recovers from a shove on ice.

The famous clip shows an engineer kicking the machine hard on a frozen slope; it scrambles, staggers and recovers. That recovery is the whole point. Dynamic balance treats disturbance rejection as the core problem rather than an exception, which is why the approach generalised to rough terrain that statically stable robots could not cross.

Raibert et al., BigDog, the rough-terrain quadruped robot

Robots in the House

200721/40

2007

ROS

An open middleware standard stops every lab rewriting drivers, transforms and message passing.

Willow Garage released the Robot Operating System as open source, giving the field a common way to pass messages between processes, manage coordinate frames and share drivers and algorithms. It is not an operating system and its real-time story took years to fix, but it made robotics software cumulative rather than perpetually restarted.

Quigley et al., ROS: an open-source Robot Operating System

Robots at Work

201022/40

2010

Aggressive quadrotor flight

Cheap MEMS sensors plus fast control make small multirotors both agile and mass-producible.

Kumar’s lab at Penn flew quadrotors through thrown hoops, perched on walls and built structures cooperatively. The enabling shift was commodity inertial sensors and brushless motors from the mobile phone and hobby supply chains, which dropped the cost of a flying platform by orders of magnitude and created the consumer drone industry alongside the research one.

Mellinger and Kumar, Minimum snap trajectory generation and control

Soft and Small

201123/40

2011

Soft robots

Whitesides builds machines from elastomer and air that squeeze through gaps a rigid robot cannot.

Pneumatic networks moulded into silicone bend when inflated, producing locomotion with no joints, bearings or motors. Soft bodies are inherently safe around people and conform to objects rather than needing exact models of them. The trade is control: a continuum body has infinite degrees of freedom and resists the usual kinematics entirely.

Shepherd et al., Multigait soft robot

Robots at Work

201224/40

2012

Amazon buys Kiva

Warehouse robots bring the shelf to the worker instead of navigating around people.

Kiva’s insight was to restructure the problem rather than solve the hard version of it. Rather than build a robot that picks items from static shelves, they built simple robots that lift entire shelving units and carry them to a stationary human picker. Hundreds coordinate through a central scheduler. Amazon paid 775 million dollars and took the system exclusive.

Wurman et al., Coordinating hundreds of cooperative autonomous vehicles

Robots at Work

201225/40

2012

Baxter and safe collaboration

A robot with series elastic actuators can share a workspace with people and be taught by hand.

Traditional industrial arms are stiff, fast and lethal, so they live in cages. Baxter used compliant actuators and force sensing so a collision is absorbed rather than transmitted, and it was programmed by physically guiding its arms. Rethink Robotics failed commercially, but collaborative robots are now a standard category and the safety standards it prompted remain.

Rethink Robotics Baxter

Soft and Small

201326/40

2013

RoboBee flies

An eighty-milligram flapping-wing robot achieves controlled flight at insect scale.

At that size conventional motors and bearings do not work, so the Harvard team used piezoelectric actuators and a folded laminate fabrication process borrowed from pop-up books. Power had to come down a tether because no battery was light enough. It demonstrated that insect-scale flight is an engineering problem, not a physical impossibility.

Ma et al., Controlled flight of a biologically inspired, insect-scale robot

Robots at Work

201527/40

2015

The DARPA Robotics Challenge

Humanoids attempt disaster response, and mostly fall over on international television.

Prompted by Fukushima, the challenge asked robots to drive, open doors, turn valves and climb stairs in a degraded environment. The blooper reel was widely shared and entirely fair: the tasks were far harder than they looked, and the falls documented exactly where the field was. It also seeded the control work behind the current generation of humanoids.

DARPA Robotics Challenge Finals

Soft and Small

201628/40

2016

Learning to grasp from data

Google runs fourteen arms for months and learns grasping from 800,000 real attempts.

Analytic grasp planning needs an accurate object model, which real clutter never provides. Instead the team let arms try, recorded what succeeded, and trained a network to predict grasp success from images. Performance improved with data alone. The arm farm made the point that robotics could be a data problem, if you were willing to pay for the data.

Levine et al., Learning hand-eye coordination for robotic grasping

Learning to Act

201629/40

2016

Atlas runs, jumps and backflips

Whole-body model-predictive control turns a humanoid from a careful walker into an athlete.

Boston Dynamics replaced the crouched, quasi-static gait of earlier humanoids with optimisation that plans contact forces over a short horizon and re-solves continuously. Atlas could leap between boxes, vault obstacles and recover from failures mid-motion. The parkour videos were choreographed, but the underlying control was genuinely general.

Boston Dynamics Atlas

Learning to Act

201830/40

2018

Domain randomisation crosses the reality gap

Policies trained only in randomised simulation transfer to physical robots without tuning.

Simulation is cheap and wrong; reality is expensive and correct. Randomising textures, lighting, friction and masses during training forces the policy to treat the real world as just another sample from the distribution. OpenAI used it to teach a hand to manipulate a cube with no real-world training at all, and sim-to-real became the default pipeline.

OpenAI, Learning dexterous in-hand manipulation

Learning to Act

201931/40

2019

Spot goes on sale

Boston Dynamics ships a legged robot as a product rather than a research platform.

Spot walks over rubble, climbs stairs and carries sensor payloads, and it is sold to anyone who signs a terms-of-use agreement forbidding weaponisation. Commercial uptake has been steady in construction progress capture, plant inspection and public safety. It is the first legged robot with a real installed base outside laboratories.

Boston Dynamics Spot

Learning to Act

201932/40

2019

Learned locomotion beats hand-tuned control

ETH Zurich trains a quadruped controller in simulation that outperforms years of analytic work.

ANYmal’s learned policy was faster, more robust and more energy-efficient than the model-based controller it replaced, and it could recover from being thrown on its back. The result was a turning point: for a problem where the classical approach was mature and well understood, reinforcement learning simply did better.

Hwangbo et al., Learning agile and dynamic motor skills for legged robots

Learning to Act

202033/40

2020

Driverless rides open to the public

Waymo removes the safety driver and starts carrying passengers in Phoenix.

Fifteen years after the Grand Challenge, a vehicle with nobody in the front seat began serving paying riders. The gap between a demo and a service was filled with remote assistance, meticulous mapping, a geofence and an operational safety case rather than any single algorithmic breakthrough. Scaling beyond mapped, benign-weather cities remains the open problem.

Waymo One

Learning to Act

202134/40

2021

Ingenuity flies on Mars

A 1.8-kilogram helicopter achieves powered flight in an atmosphere one percent as dense as Earth’s.

Rotors had to spin above 2,500 rpm in near-vacuum, and the round-trip light delay meant every flight was fully autonomous. Planned for five demonstration flights, Ingenuity completed 72 over nearly three years and became a scout for Perseverance. It proved aerial mobility works on other planets, which changes what future missions can reach.

NASA Ingenuity Mars Helicopter

Foundation Models for Robots

202135/40

2021

Robots in the field

Vision-guided machines weed and harvest crops, reducing herbicide use by an order of magnitude.

Agricultural robotics works because the environment is structured, the task is repetitive and labour is genuinely scarce. Systems classify crop against weed per plant and act on each individually, whether with a targeted spray, a blade or a laser. It is one of the few domains where field-deployed autonomy is already unambiguously economic.

Carbon Robotics LaserWeeder

Foundation Models for Robots

202336/40

2023

Vision-language-action models

RT-2 fine-tunes a web-scale vision-language model to output robot actions directly.

Instead of training control from robot data alone, RT-2 starts from a model that already knows what a banana is and that rubbish goes in bins, then teaches it to emit action tokens. The robot could follow instructions referring to objects and concepts never present in its own training, because the semantics came from the web.

Brohan et al., RT-2: Vision-language-action models

Foundation Models for Robots

202337/40

2023

Open X-Embodiment

Thirty-four labs pool data from 22 robot types, and a single policy transfers across all of them.

Robot learning had been fragmented because every lab used different hardware and no dataset was large enough. Pooling more than a million trajectories produced models that outperformed the specialists on their own robots. It was robotics’ ImageNet moment: the demonstration that shared data beats bespoke data, even across different bodies.

Open X-Embodiment Collaboration

Foundation Models for Robots

202438/40

2024

Humanoids enter pilot production

Figure, Tesla, Agility and Unitree put general-purpose humanoids into real workplaces.

The bet is that a human-shaped robot needs no changes to buildings, tools or processes, so one platform can amortise across many tasks. Pilots are running in warehouses and automotive plants at small scale. Whether the economics work is genuinely unresolved: the humanoid form is the hardest way to build a machine and the easiest way to deploy one.

Agility Robotics Digit

Foundation Models for Robots

202439/40

2024

Teleoperation as a data pipeline

Low-cost puppeteering rigs let humans generate the demonstration data robot learning is starved of.

Text and images could be scraped; robot trajectories cannot. Projects like ALOHA and its open-source successors cut the cost of a bimanual teleoperation setup to a few thousand dollars, so demonstrations can be collected in volume by non-specialists. The bottleneck in robot learning is now a logistics problem about human hours, which is a very different kind of problem.

Zhao et al., Learning fine-grained bimanual manipulation with low-cost hardware

Foundation Models for Robots

202640/40

2026

Moravec’s paradox, still

Reasoning got cheap and manipulation did not.

Moravec observed in 1988 that the hard problems for machines are the ones evolution solved long ago: perception and dexterity, not chess. Four decades on, a model can pass professional exams while no robot reliably folds laundry. Data, not algorithms, is now the binding constraint — and unlike text, robot data cannot be scraped.

Moravec, Mind Children

End of the line

40 checkpoints, and laundry is still unsolved.

The through-line is Moravec’s paradox. Every era assumed the remaining problems were cognitive, and every era discovered they were physical. The body turned out to be the hard part.

Back to the site Read the essays

1898Scroll or drag