“Rebooting AI” notes-5
(a series of posts for each chapter)
<<< Previous post (“Rebooting AI” notes-4)
>>> Next post (“Rebooting AI” notes-6)
In this post I would like to list some quotes from the book which I found them either important or interesting. (Note: some parts of the quotes are paraphrased by myself and some of them are my own insights)
5. Where’s Rosie?
Are you worried about superintelligence robots rising up and attacking us? Don’t be. At least for now, here are six things you can do in the event of a robot attack:
- Close your doors or lock them: contemporary robots struggle greatly with doorknobs, i.e. they can not generalize to all different kinds of doorknobs yet.
- Paint your doorknob black, against a black background: will greatly reduce the chance that the robot will even be able to see it, i.e. color is one of the most difficult tasks for CV (computer vision) algorithms.
- Put a big poster of a school bus on your front door or wear a t-shirt with a picture of a cute baby: the robot will be totally confused, think you are a baby, and leave you alone (considering the robot is caring one).
- If that doesn’t work, go upstairs, and leave a trail of banana peels and nails along the way: few robots be able to handle an impromptu obstacle course.
5. Hop up on the table: even the ones that climb stairs probably can’t hop up onto a table.
6. Call for the help: robot’s battery will soon run out. Free-ranging robots currently typically last for a few hours, but not much more between charges, because the computers inside them demand such vast amounts of energy.
Any intelligent creature needs to compute five basic things: where it is, what is happening in the world around it, what it should do right now, how it should implement its plan and what it should plan to do over the longer term in order to achieve the goals it has been given. The field of robotics has gotten pretty good at implementing some parts of the robot’s cognitive cycle. For instance, localization and motor control.
Localization: is harder than it might look like. The obvious way to start is with GPS. GPS may not be that much accurate all the time (especially indoors). Fortunately, robots can leverage a few complementary features to make its localization more accurate: dead reckoning (which track’s a robot’s wheels to estimate how far it has gone), vision (scene recognition), and maps. Roboticists have also developed a family of techniques called SLAM (Simultaneous Localization And Mapping) which allows robots to put together a map of their environment and to keep track of where they are in the map and where they are headed. At each step, within the SLAM algorithm, the robot goes through the following steps:
- uses its sensors to see the part of its environment that is visible from its current position
- improves is current estimate of its position and orientation by matching what it is seeing against objects in its mental map
- add to its mental map any objects, or parts of objects, that it has not seen before
- either moves or turns, and adjusts its estimate of its new position and orientation by taking into account how much it has moved or turned.
Motor control: the job of guiding a robot’s motions, such as walking, lifting things, rotating its hands, turning its head, or climbing stairs. Suppose that there is a cup of tea on a table, and a humanoid robot is supposed to reach out its arm and grasp the cup’s handle between two fingers. Robot has to figure out how to move the various parts of its arm and hand so that they end up at the right place without, in between, bumping into the table, having one part bumping into another, or knocking the teacup over. Then it has to exert enough force on the teacup handle that it has a firm grasp but not so much force that it shatters the china. As you may imagine, these steps need great control mechanism. Despite the complexity of the problem, in recent years there has been considerable progress. Especially Boston Dynamics (the robot company) run by Marc Raibert (a researcher with deep learning in the problem of human and animal motor control) has released many robots with superior motor control such as BigDog, SpotMini, etc.
