Moravec's paradox says that reasoning tasks that seem hard to people are fairly easy to program into computers, while tasks that come easily to us, like moving around, seeing or grasping things, are quite hard to program. Picking things up is hard for robots: a robot has trouble picking up an actual chess piece, even though a computer can beat a chess grandmaster. Many factory robots are programmed to do the same action again and again, the same way each time. Most robots are programmed in one of two ways. In learning from demonstration, they watch a task being done and copy it, whereas in motion planning a programmer spells out the task's goals and limits. A robot that learns a skill by watching often cannot easily use that skill in a new situation and stay accurate.

At the 2015 DARPA Robotics Challenge, a humanoid robot nicknamed Running Man fell flat on its back instead of stepping onto a small pile of cinder blocks. A robotics researcher says walking on two legs is hard for robots: no robot with two legs has yet matched the control, toughness and efficiency of people, or even chickens. People have been walking on two legs for hundreds of thousands of years, and birds such as ostriches for millions of years. Many humanoid robots walk in a careful, conservative way, unlike the more dynamic walking of people. A robot that walks on legs can fall over, and a hard enough push can topple it. For a walking robot, small errors in where a foot lands matter, since being off by a few inches can leave the robot staggering.

One way to plan a walking robot's steps is called capture point: the spot on the ground that a robot's foot must reach to stop its body from falling over. At the DARPA Robotics Challenge, one robot had wheels on its knees and feet, so it could roll in a stable kneeling position through much of the course. One of the most important jobs robots do is to protect people from injury by doing dangerous jobs in their place. The 2011 Fukushima nuclear plant disaster might have been lessened if robots had been able to drive into the plant, get past stairs or debris, and turn valves or switches.

In 2016 Harvard scientists built the octobot, a small 3D-printed robot with a body that is entirely soft. The soft robot was nicknamed the octobot, and octopuses have long been a source of inspiration in soft robotics. Soft robots are hard to make because batteries and circuit boards are rigid, so until the octobot, soft robots were either tied to an outside power system or had hard parts. A chemical reaction inside the octobot turns a little liquid fuel into a lot of gas, which flows into its arms and inflates them like a balloon. The octobot has no electronics: tiny channels holding chemical reactions control its movements, so it needs no battery or circuit board. Its gas flow was timed by a soft circuit that works like a simple electronic oscillator, and the work was reported in the journal Nature.

Robot designers have tried other bodies too. In 2014 Harvard scientists made 1,024 tiny robots, called Kilobots, that could build shapes together as a swarm. Before the Kilobots, most robot swarms had fewer than 100 robots. Each Kilobot was extremely simple, just a few centimeters across and standing on three pin-like legs. The Kilobots had no robot or person in charge of the shape: each followed simple rules. The researchers took ideas from nature: just as trillions of cells can form one body, or thousands of starlings can form a flowing flock, many simple robots can make something complex. Computer scientists see the Kilobot swarm as a milestone for collective artificial intelligence, in which many simple units work together.

Words to know

• paradox: puzzling idea
• humanoid: person-shaped
• conservative: careful
• octobot: soft robot
• grandmaster: top chess player