A short history of robots and embodied AI

Robots are older than most people think, and the history explains the shape of the field. The classical toolkit (kinematics, control, planning) came from decades of engineers who had to make machines work without machine learning. The recent wave is not a replacement for that toolkit. It sits on top of it.

The word and the dream#

The word "robot" comes from the 1920 play R.U.R. by the Czech writer Karel Čapek, from the Czech robota, meaning forced labour (Čapek credited his brother Josef with coining the word). The idea of a made worker is far older, but the play gave it the name. For decades robots lived in fiction while engineers worked on pieces of the problem: servo motors, feedback control and early computers.

Era 1: machines that repeat (1950s to 1980s)#

The first useful industrial robot, Unimate, went to work at a General Motors plant in 1961, lifting hot die-cast parts. George Devol had filed a patent for the idea in 1954 (granted in 1961). These machines had no sensors worth the name. They replayed a recorded sequence of joint positions, again and again, with great repeatability.

That is still the foundation of factory automation: if the world is arranged to be exactly the same every time, a blind, precise arm is the cheapest and most reliable answer. The cost of this era's approach is rigidity. Change the part by a few millimetres and the robot fails.

Era 2: robots that reason (1960s to 1980s)#

In parallel, researchers asked how a machine might think. Shakey, built at SRI in the late 1960s, was a wheeled robot with a camera and a range finder that planned how to move boxes between rooms. The work behind it produced lasting tools, including the A\* search algorithm (chapter 49) and the STRIPS planner. Shakey was slow and fragile, and it taught the field a lesson that still holds: reasoning about a clean model of the world is easy, and connecting it to the messy real one is hard.

Era 3: reactive and probabilistic robots (1980s to 2000s)#

Two shifts changed mobile robots.

• Behaviour-based robotics. In 1986 Rodney Brooks argued for layers of simple reactive behaviours instead of one big world model. The robot reacts quickly to what it senses. Brooks co-founded iRobot, whose Roomba (2002) became a mass-market home robot, selling a million units by 2004.
• The probabilistic turn. Sensors lie. Researchers such as Sebastian Thrun built the maths of uncertainty into robotics: Kalman filters, particle filters and SLAM (chapters 41 to 44). Stanford's car Stanley won the 2005 DARPA Grand Challenge, a 132-mile desert race, using these methods.

The DARPA challenges (2004, 2005 and the 2007 Urban Challenge) helped launch the self-driving industry. Around the same time Honda's ASIMO (2000) showed a humanoid walking on stage, an engineering marvel with little autonomy.

Era 4: robots that learn (2010s)#

Deep learning transformed perception from 2012: object detection, segmentation and depth estimation became reliable. Then learning began to reach control:

• Reinforcement learning in simulation. In 2019 researchers at ETH Zürich trained the quadruped ANYmal to get up and walk in simulation and then ran the controller on the real robot. OpenAI's Dactyl hand learned to reorient an object in its fingers (2018), and later to solve a Rubik's cube (2019), the same way, with heavy randomisation of the simulated world (chapter 65).
• Learning from demonstration. Policies trained on a few hundred human demonstrations could do tasks that had resisted hand-coding (chapter 66).

Meanwhile the hardware changed too. Smartphones made cameras and inertial sensors cheap, electric cars made batteries and motor controllers cheap, and companies such as Boston Dynamics showed what dynamic legged machines could do.

Era 5: robot foundation models (2022 onward)#

The language-model recipe (a large transformer trained on broad data) arrived in robotics:

These are vision-language-action models (chapter 76). They take camera images and an instruction in words and output motor commands. They generalise better than earlier policies, but they are far from solved: reliability, speed and safety are open problems (chapter 83).

What the history teaches#

• Each era kept the last one's tools. Modern robots still use inverse kinematics, PID loops and Kalman filters under their learned parts.
• The bottleneck moves. First it was mechanics, then computing, then perception, and now data and reliability.
• Demos run ahead of deployment. A robot that works nine times in ten is a great demo and an unusable product. The gap between the two is where most of the work in robotics happens.