The moral of the story is, like everything in AI, machine learning, robotics, cybernetics etc., no matter how complicated you think it is, it's more complicated than that.
Nope...sometimes it's way simpler: hey maybe we can just throw tons of data at a nueral net is something few would have believed would work 15 years ago.
I kidnapped my robot and held it hostage in the kitchen. It took about ten minutes, but it found its way back to the dock in the hallway. It repeatedly stopped and did a 360 degree scan, before eventually figuring out where home was. Not sure if it was by calculation or luck. It certainly gave the impression of stopping to "have a think about it." Still, impressive for a four year old machine.
I’m glad she called out the quality of those robots. We were given these robots and told to localize them for the first design class at McGill in software engineering, and we had so many issues with drift. Really good video! I found the presentation of the ideas to be quite clear and effective
Remember when this was applicable to daily life? After a night out, trying to fit keys into a door, constantly updating your belief that the keys aren't your bike keys, and the door is your own front door, and both your sensors and actuators were severely hampered by voluntary ethanol intoxication...
I love how they focus on "imperfect world" of robotcs, such as part of robot not working 100% perfectly. It's kind of difference between software and hardware engineer, where software engineer knows that with very high reliability what it asks from program will happen exactly as he told it to. With hardware you have to go with many imperfections and tolerances because real world is.... well not perfect :)
14:50 - sounds a lot like what I do when getting to know a new city... I’ll build mental maps of the different places I go (often centered around transit stations or the like, and eventually, as I’m expanding out a map from one location, it’ll suddenly connect to a map from a different location, and then I’ll link them and have a bigger map with multiple starting locations contained within it. :) Nice video! Thanks Ayse! I hope we’ll indeed hear more from you, about SLAM or whatever else!
SLAM - Simultaneous Localization and Mapping! I love the design of that two-wheeled robot. I think I will build a replica before the next video comes out so I can follow along at home! I've implemented SLAM on the adorable Turtlebot 3 Burgers using spinny LiDaR so I think I can hack something together with my old Lego kits! So excited to see more robots on the channel!!
We have your robot. As proof, we are enclosing one of its sensors. We will send you another part of your robot every day until there are no parts left. If you ever want to see your robot again, you must send us a mint condition Lego Deathstar kit.
No problem, you just wait until they send all parts so you can simply reassemble the robot :P. That's the great difference between machines and humans: If someone sends you "parts" of a human body you "usually" can't reassemble him.
Sorry, but I have to say this: 11:41 into the clip: "I knew I was here". I can't accept that. It can't know it was there. It is still trying to determine where it is. This is enforced by the line being called BELIEF. It isn't a certainty.
Literally, a textbook example from "Probabilistic Robotics" by Thrun, Burghard, Fox :) Even though this is from 2005, I would still recommend this, if you want to really learn the essentials of SLAM.
The robot knows where it is at all times. It knows this because it knows where it isn't, by subtracting where it is, from where it isn't, or where it isn't, from where it is, whichever is greater, it obtains a difference, or deviation. The guidance sub-system uses deviations to generate corrective commands to drive the robot from a position where it is, to a position where it isn't, and arriving at a position where it wasn't, it now is. Consequently, the position where it is, is now the position that it wasn't, and it follows that the position where it was, is now the position that it isn't. In the event of the position that it is in is not the position that it wasn't, the system has required a variation. The variation being the difference between where the robot is, and where it wasn't. If variation is considered to be a significant factor, it too, may be corrected by the GEA. However, the robot must also know where it was. The robot guidance scenario works as follows: Because a variation has modified some of the information the robot has obtained, it is not sure just where it is, however it is sure where it isn't, within reason, and it knows where it was. It now subtracts where it should be, from where it wasn't, or vice versa. By differentiating this from the algebraic sum of where it shouldn't be, and where it was. It is able to obtain a deviation, and a variation, which can be used to correct itself.
Could putting a "brush sensor" at the bottom of the robot to check for movements help? Maybe record the rotation and movement only at the moments when the brush senses some movement? Also maybe a gyroscope and some balls that move around for when picked up, or the thing that watches/ phones use to record steps. Maybe
I think the perfect example of a device which compensate many things in the space is a drone! Try to fly it without sensors! It wants to fly in any direction possible but with gyro etc. and some logics it can actually fly quite well!
The way Roombas handle this is hilarious and extremely effective. They just drive around randomly and try to measure the distance from themselves to the base station. It already drivers essentially randomly to operate normally anyway, so it essentially just keeps doing that. More often than not, it will pick up the base station and can dock when the battery gets low. maybe 1 in 10 times it will not find the base station before it dies. And it only has one distance sensor! its pretty genius for how simple it is
11:02 Great point that can be projected on us humans who also have the power to move and gather more information to subsequently change or shift our believes in a certain matter. Still some people choose not to move and insist on their opinion from their static point of view
Or they move, but ignore certain things which for others are clear features of the landscape... Sensor fusion between diverse channels of information is the key!
Where am I? Environmental senses and some algorithms can answer that. How am I? Internal sensors and algorithms can answer that. Why am I? A random number generator and a some algorithms can answer that.
It reminds that every time I visit my inlaws, I go to their neighbor and spin their cheap lawnmover robot 180 degrees - and they must reset its guidance-system. (Yes, I've been 9 y/o for about 50 years now) ;-)
Read Markovs Localization, Kalman Filters and Particle filters to learn more about localization If you are interested in Robotics, here is a brief overview: - Perception (Figuring out features and landmarks in your environment, and their representation) - Localization (Figuring out, where are you located, based on current perception) - Planning (Planning a set of actions to reach your goals) - Control (Sending actual control commands to the robot motors, to move the robot to the planned location) All of these steps have some form of noise, hence probabilistic methods are used to take conservative actions, which allow you to move through the environment with minimum risk of collision/failure
need proximity sensor too create and 3d map mesh from loggin current environment, then navigate using coordinates, in a loop spiral from centre point, then it has chance to learn surroundings. my opinion
Didn't that plane crash, (partially) because it didn't properly fuse all its sensors that provide the altitude? It only trusted one sensor (that failed) even though other sensors had conflicting measurements?
The robot knows where it is at all times. It knows this because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is (whichever is greater), it obtains a difference, or deviation. The guidance subsystem uses deviations to generate corrective commands to drive the missile from a position where it is to a position where it isn't, and arriving at a position where it wasn't, it now is.
I've seen people locate a location on Earth through an image by looking things up on Google maps / images. Maybe a robot could do something similar with a local database or sat connection and gps? For inside location it would be more difficult.
If the robot moves forward, surely the probability peaks move *backwards* !? (I can not see that they can be in spacial reference to anything else than the robot itself.)
No, having position data relative to the robot would mean that the robot is "right here" at all times. That wouldn't be useful to know. However, if the robot builds a map, all the things in the map move backwards with respect to the robot when it moves forwards.
Is there a reason we're pretending we can't just put a 15 cent GPS chip in the robot? Or how about even a compass that logs movement relative to magnetic north? Or what about relying on the RF environment rather than visual for spaces such as inside a home (assuming LiDAR is still just too expensive for distancing mapping entire room) and triangulating based on wifi routers, pings from different devices, etc? Sure devices might move around, but it's probably safe to assume that they haven't ALL moved at once, and should be not terribly difficult to triangulate their relative positions from one another and orient ones 'self' in relation. It just seems like using visual sensors would be one of the last things I'd reach for given the very different capacities robots have for gathering sensory data compared to humans. Short of LiDAR, using simple RADAR with ultrasonic pings would be good getting room dimensions and relative distance to each wall, wouldn't it? I can see difficulties if you want to support total discontinuities in sensory input (powering device off, moving it, turning it back on) but I'd expect in operation that could mostly be avoided with batteries and such.
Having said this to a later comment: Unfortunately GPS seldom solves the localisation problem: 1) mostly unavailable in indoor (and remote outdoor) environment, 2) has meter-level accuracy, which is just not good enough for localisation. But of course it is a good source of information to embed in your algorithm where available to improve predictions.
Most robot kidnappers I know have deep, underground lairs or heavily shielded hideaways. Ok, so they live in their parent's basement and covered the walls and ceiling with tin foil... times are hard! 😁
So a kidnapped robot is in a room it has never seen before. It doesn't know where the room is in the world. It turns on GPS and gets LAT/LONG data. What corner of the room is the robot near?
This is frustrating as hell. I would love to discuss military guidance systems but thanks to ITAR the chance of that happening is zero. Great video, you guys are on the right track. It's the first I've heard someone use the word propreoception in conversation, I first came across the concept a few decades ago in a book called "The Man Who Mistook His Wife For A Hat" by Oliver Sacks which is a great read if anyone is interested in psychological and physiological conditions in humans. Peace.
3D printer setups "just level everything, a good 3d printer should not need sensors!!!" Real people using 3d printers [fits bump sensors and autolevelling devices].
just this weak I needed to implement a particle filter on a robot (no joking). I hadn't need to implement one for years, and the same weak, a computerphile video about it.
Agreed. … And to give some sample boundary conditions on scope and precision, e.g.: what the robot knows /a priori/, and if it is to locate itself on this particular desk, in this gym, or anywhere on Earth? GPS makes its overdue appearance at about 90% in.
![The moment we stopped understanding AI [AlexNet]](/img/1.gif)
The moral of the story is, like everything in AI, machine learning, robotics, cybernetics etc., no matter how complicated you think it is, it's more complicated than that.
Yep years of effort lead to this moment. Just for the robot to hit the walls of a room trying to find where it is
Nope...sometimes it's way simpler: hey maybe we can just throw tons of data at a nueral net is something few would have believed would work 15 years ago.
Yes ! Machine learning , natural language processing , is already complicated , and it application (chatbot,image recognition) is more complicated !
But also that scientists and engineers have always better answers to those problems than you'd think once you learned how complicated they are.
I kidnapped my robot and held it hostage in the kitchen. It took about ten minutes, but it found its way back to the dock in the hallway. It repeatedly stopped and did a 360 degree scan, before eventually figuring out where home was. Not sure if it was by calculation or luck. It certainly gave the impression of stopping to "have a think about it."
Still, impressive for a four year old machine.
Computerphile: kidnapped robot problem
My roomba at home: hm, yes, interesting...
9:17... i see it everywhere i go i cannot escape
OMG ! SHE WAS MY TEACHER IN TURKEY YEDITEPE UNI FOR COMP. ENG. PROGRAMME. she is explaining the subject with enthousiasm as always.
I’m glad she called out the quality of those robots. We were given these robots and told to localize them for the first design class at McGill in software engineering, and we had so many issues with drift. Really good video! I found the presentation of the ideas to be quite clear and effective
As part of my research, I am working on designing acoustic SLAM so that through echolocation we can construct a map of an environment
It's been so long since we had a live meetup interview... this is so great!
Her enthousiasm makes me want to study robotics...
we need a robotphile channel
She's so clear and fun! Would love to have more content with her!
whenever I pick up my roomba: _where did you put me, human?!_
SLAM algorithm would be a great topic (or series of topics).
Remember when this was applicable to daily life? After a night out, trying to fit keys into a door, constantly updating your belief that the keys aren't your bike keys, and the door is your own front door, and both your sensors and actuators were severely hampered by voluntary ethanol intoxication...
I love how they focus on "imperfect world" of robotcs, such as part of robot not working 100% perfectly. It's kind of difference between software and hardware engineer, where software engineer knows that with very high reliability what it asks from program will happen exactly as he told it to. With hardware you have to go with many imperfections and tolerances because real world is.... well not perfect :)
14:50 - sounds a lot like what I do when getting to know a new city... I’ll build mental maps of the different places I go (often centered around transit stations or the like, and eventually, as I’m expanding out a map from one location, it’ll suddenly connect to a map from a different location, and then I’ll link them and have a bigger map with multiple starting locations contained within it. :)
Nice video! Thanks Ayse! I hope we’ll indeed hear more from you, about SLAM or whatever else!
Haha yep I’ve made that comparison a lot before! It’s very similar to the intuition of SLAM/loop closures
I find it so cool that we can map our every-day thought processes into mathematical formulas. Thanks again for the great video!
Just pay the ransom. 💶 ------> 'Kidnapped Robot Problem' solved 🤖
We do not negotiate with filthy botnappers!! ☝️🧐
I didn't expect to see a LEGO Mindstorms robot in this video
SLAM - Simultaneous Localization and Mapping! I love the design of that two-wheeled robot. I think I will build a replica before the next video comes out so I can follow along at home!
I've implemented SLAM on the adorable Turtlebot 3 Burgers using spinny LiDaR so I think I can hack something together with my old Lego kits!
So excited to see more robots on the channel!!
Ayse can play a mean game of Pictonary! Greetings from LCAS!
Inside a jawa sand crawler if you're lucky, or the spice mines of Kessel if you're not.
So ... Confuse the Terminator by giving it 3 doors... gotcha!
We have your robot. As proof, we are enclosing one of its sensors. We will send you another part of your robot every day until there are no parts left. If you ever want to see your robot again, you must send us a mint condition Lego Deathstar kit.
No problem, you just wait until they send all parts so you can simply reassemble the robot :P. That's the great difference between machines and humans: If someone sends you "parts" of a human body you "usually" can't reassemble him.
She is incredible and elegant in her way of explaining. Would love to see more of her!
This sounds like a kalman filter, I tried to use it some time ago with my non university and non maths background, that was quite hard tbh
Beautiful, meticulous editing, as always.
This is propably the cutest robot drawing i have ever seen :D
I came here to say this.
amogus robot (i am sorry)
I've immediately noticed the Turkish accent. Greetings from Turkey and good luck to you ransoming your robot 😅
Me too :)
Sounds to me based on her accents like she is Turkish and living in the UK.
Instructions unclear: I now have a DARPA's robot in my basement!
Sorry, but I have to say this: 11:41 into the clip: "I knew I was here".
I can't accept that. It can't know it was there.
It is still trying to determine where it is.
This is enforced by the line being called BELIEF. It isn't a certainty.
Literally, a textbook example from "Probabilistic Robotics" by Thrun, Burghard, Fox :)
Even though this is from 2005, I would still recommend this, if you want to really learn the essentials of SLAM.
The three door hallway problem? I think I used the same textbook in school too!
Really great explanations. Robotics is super interesting!
The robot knows where it is at all times. It knows this because it knows where it isn't, by subtracting where it is, from where it isn't, or where it isn't, from where it is, whichever is greater, it obtains a difference, or deviation.
The guidance sub-system uses deviations to generate corrective commands to drive the robot from a position where it is, to a position where it isn't, and arriving at a position where it wasn't, it now is. Consequently, the position where it is, is now the position that it wasn't, and it follows that the position where it was, is now the position that it isn't. In the event of the position that it is in is not the position that it wasn't, the system has required a variation. The variation being the difference between where the robot is, and where it wasn't. If variation is considered to be a significant factor, it too, may be corrected by the GEA. However, the robot must also know where it was. The robot guidance scenario works as follows: Because a variation has modified some of the information the robot has obtained, it is not sure just where it is, however it is sure where it isn't, within reason, and it knows where it was. It now subtracts where it should be, from where it wasn't, or vice versa. By differentiating this from the algebraic sum of where it shouldn't be, and where it was. It is able to obtain a deviation, and a variation, which can be used to correct itself.
Always curious how my Roomba knew where it was when I move it after it gets stuck. Great presenter, would love to watch her explain SLAM.
Could putting a "brush sensor" at the bottom of the robot to check for movements help?
Maybe record the rotation and movement only at the moments when the brush senses some movement?
Also maybe a gyroscope and some balls that move around for when picked up, or the thing that watches/ phones use to record steps.
Maybe
I think the perfect example of a device which compensate many things in the space is a drone! Try to fly it without sensors! It wants to fly in any direction possible but with gyro etc. and some logics it can actually fly quite well!
I'm starting to think about helping that poor robot to escape.
Poor robot, just trying to find out where it is ):
poor guy can't even contemplate who he is.
if only humans would have this approach about their believes. constantly updating the probability of assumed facts. i like that.
The way Roombas handle this is hilarious and extremely effective. They just drive around randomly and try to measure the distance from themselves to the base station. It already drivers essentially randomly to operate normally anyway, so it essentially just keeps doing that. More often than not, it will pick up the base station and can dock when the battery gets low. maybe 1 in 10 times it will not find the base station before it dies. And it only has one distance sensor! its pretty genius for how simple it is
11:02 Great point that can be projected on us humans who also have the power to move and gather more information to subsequently change or shift our believes in a certain matter.
Still some people choose not to move and insist on their opinion from their static point of view
Or they move, but ignore certain things which for others are clear features of the landscape... Sensor fusion between diverse channels of information is the key!
So, the robot is an empiricist, not a racionalist... And a strong bayesian?
Please, make a video about SLAM's
Where am I?
Environmental senses and some algorithms can answer that.
How am I?
Internal sensors and algorithms can answer that.
Why am I?
A random number generator and a some algorithms can answer that.
I've been tryna kidnap a rtx 3090 for a while now
9:20 did she just draw an among us person?
i was about to say that
I was about to say that :D
ඞ
ඞ 🚪 🚪 🚪
also looks a bit like Bender's head on wheels 🤖
Could you please put some subtitles on, maybe just auto generated
It reminds that every time I visit my inlaws, I go to their neighbor and spin their cheap lawnmover robot 180 degrees - and they must reset its guidance-system. (Yes, I've been 9 y/o for about 50 years now) ;-)
What are these propabalistic methods called, wich are used for localisation?
Very interesting video
Read Markovs Localization, Kalman Filters and Particle filters to learn more about localization
If you are interested in Robotics, here is a brief overview:
- Perception (Figuring out features and landmarks in your environment, and their representation)
- Localization (Figuring out, where are you located, based on current perception)
- Planning (Planning a set of actions to reach your goals)
- Control (Sending actual control commands to the robot motors, to move the robot to the planned location)
All of these steps have some form of noise, hence probabilistic methods are used to take conservative actions, which allow you to move through the environment with minimum risk of collision/failure
Please do the SLAM algorithm aswell
I have had the same problem most of my life.
When she drew the three doors and brought in probability and belief I was sure Monty Hall would make an appearance.
Perfect explanation :)
Was there a ransom note?
really fascinating subject
Interesting content and very engagingly explained 👍
need proximity sensor too create and 3d map mesh from loggin current environment, then navigate using coordinates, in a loop spiral from centre point, then it has chance to learn surroundings. my opinion
Different angles, different kidnappers. Same world.
"this is a toy"
... In your opinion! Download some more ram and feed it after midnight and call it a toy again!
If robots don't know where to go then they should ask for directions.
If a robot doesn't care where it is, is it really kidnapped?
watch the next video on Philosophile channel :D
Add subtitles
Didn't this happen to Lt. Cmd Data?
Now I’m imagining a roomba bumbling around in the Ashtray Maze from Control. 😂
Please create a video on SLAM too.
Discussed in the video.
Thanks 😊
If you want to draw a decent square, you draw 2 paralel vertical lines and 2 parallel horizontal lines.
I did wonder whether you could disable spot with a bedsheet
There's a very simple answer to this problem - GPS! (OK, the resolution isn't actually good enough for that...)
Why the reup?
There was a problem with the footage
Wow, that guy in the back just doesn't care that you are doing a recording.
The cybermen don't care about the business of puny humans
And then I got to the point in the video where there actually was a guy in the background...
Didn't that plane crash, (partially) because it didn't properly fuse all its sensors that provide the altitude? It only trusted one sensor (that failed) even though other sensors had conflicting measurements?
The robot stares on even after being called "Cheap" at 3:34
@Computerphile can you do a video about the 54 years old bug in the Universal Turing Machine?
The robot knows where it is at all times. It knows this because it knows where it isn't. By subtracting where it is from where it isn't, or where it isn't from where it is (whichever is greater), it obtains a difference, or deviation. The guidance subsystem uses deviations to generate corrective commands to drive the missile from a position where it is to a position where it isn't, and arriving at a position where it wasn't, it now is.
she is trying so hard to avoid the word SLAM
I've seen people locate a location on Earth through an image by looking things up on Google maps / images.
Maybe a robot could do something similar with a local database or sat connection and gps?
For inside location it would be more difficult.
Stuff Made Here should probably watch this
He probably did, for what he had built, he's always ahead of pretty much everybody...
If the robot moves forward, surely the probability peaks move *backwards* !? (I can not see that they can be in spacial reference to anything else than the robot itself.)
No, having position data relative to the robot would mean that the robot is "right here" at all times. That wouldn't be useful to know.
However, if the robot builds a map, all the things in the map move backwards with respect to the robot when it moves forwards.
just imagining a good gps based robot and scrabbling the signal to make a robot not know where its at so it assumes it kidnapped
Is there a reason we're pretending we can't just put a 15 cent GPS chip in the robot? Or how about even a compass that logs movement relative to magnetic north? Or what about relying on the RF environment rather than visual for spaces such as inside a home (assuming LiDAR is still just too expensive for distancing mapping entire room) and triangulating based on wifi routers, pings from different devices, etc? Sure devices might move around, but it's probably safe to assume that they haven't ALL moved at once, and should be not terribly difficult to triangulate their relative positions from one another and orient ones 'self' in relation. It just seems like using visual sensors would be one of the last things I'd reach for given the very different capacities robots have for gathering sensory data compared to humans. Short of LiDAR, using simple RADAR with ultrasonic pings would be good getting room dimensions and relative distance to each wall, wouldn't it? I can see difficulties if you want to support total discontinuities in sensory input (powering device off, moving it, turning it back on) but I'd expect in operation that could mostly be avoided with batteries and such.
The noise in the background is really annoying.
Next can we get a bunch of robots in to the game from the movie saw?
Robot kinda plays civilization
Amazing
I'm more interested in the liquid metal time traveling homicidal robot problem.
1. GPS.
Done.
Having said this to a later comment: Unfortunately GPS seldom solves the localisation problem: 1) mostly unavailable in indoor (and remote outdoor) environment, 2) has meter-level accuracy, which is just not good enough for localisation. But of course it is a good source of information to embed in your algorithm where available to improve predictions.
Most robot kidnappers I know have deep, underground lairs or heavily shielded hideaways. Ok, so they live in their parent's basement and covered the walls and ceiling with tin foil... times are hard! 😁
So a kidnapped robot is in a room it has never seen before. It doesn't know where the room is in the world. It turns on GPS and gets LAT/LONG data. What corner of the room is the robot near?
When subtitles?
this is the area where humans are better than robots... (just a realistic point of view)
6:53 or Mars....
Mooooooore robots!
I've taken a few robots in my day.
This is frustrating as hell. I would love to discuss military guidance systems but thanks to ITAR the chance of that happening is zero.
Great video, you guys are on the right track. It's the first I've heard someone use the word propreoception in conversation, I first came across the concept a few decades ago in a book called "The Man Who Mistook His Wife For A Hat" by Oliver Sacks which is a great read if anyone is interested in psychological and physiological conditions in humans.
Peace.
3D printer setups "just level everything, a good 3d printer should not need sensors!!!"
Real people using 3d printers [fits bump sensors and autolevelling devices].
I was hoping this video had a math formula. 🙁
Noice time to kidnap some robots!
just this weak I needed to implement a particle filter on a robot (no joking). I hadn't need to implement one for years, and the same weak, a computerphile video about it.
hi
wish she started with defining what self localization means...
I think it just means finding where you are yourself! I have to do this every time I wake up...
finding out where it is by itself
Agreed. … And to give some sample boundary conditions on scope and precision, e.g.: what the robot knows /a priori/, and if it is to locate itself on this particular desk, in this gym, or anywhere on Earth?
GPS makes its overdue appearance at about 90% in.
😃😃😃😃