Agile soccer skills for a bipedal robot with deep reinforcement learning
twitter.com
twitter.com
US + Allies had clearly superior technological force, but still "lost" to some pretty irregular forces.
I don't think it is so black and white. If Russia/china/pick-your-bogey-man-of-choice invaded your country tomorrow with a robot army and started taking away your way of life (eradication of your native language, your native currency, your media, history, certain foods, schools etc), how do you think that would go down? You'd just roll over and accept everythingyou know and love about where you live being permanently removed because their robots were better? It is the same reason that we can't just have a game of chess to decide it all: when it comes down to trying to force someone to do something against their will, the ultimate and final fall back position is physical force and so sadly violence one way or another.
The military did its job.
If a country can’t stand on its own, the military can’t fix that.
“There were 2,402 United States military deaths in the War in Afghanistan (2001–2021). 1,921 of these deaths were the result of hostile action.”
https://en.m.wikipedia.org/wiki/United_States_military_casua....
How many organized people were they fighting in Iraq or Afghanistan? With what sort of equipment/regimentality/strategic coordination/etc.
As horrible as the Afghanistan war was, it would have been even worse if they had just carpet bombed everything. Add robotics, and it quickly enters nightmare fuel worse than the Terminator movies.
Both the US and the Soviets lost because Afghanistan isn't a country in a traditional sense; it's just an area of land with _lots_ of regions countrolled by war lords that doesn't have allegiance to anything central (i.e. "the country Afghanistan").
It's impossible to win a war against something like that, unless - as you correctly points out - you obliterate everything.
Winning means you installed/recovered democracy.
Had they "obliterated" everything then the US + Allies (or what allies they might have still got left after said obliteration) would have lost even more on the grand scale of things compared to a remote war in Central Asia. No amount of Hollywood + Western Media white-washing would have managed to get that stink off.
Hasn't the US carpet bombed Vietnam and still lost?
That was the issue with Vietnam (and with Afghanistan in a certain way)
The Soviet Union did that and also lost. The only way to win in Afghanistan is by fighting in Pakistan. And that means a full invasion, not a half-measure as was tried and failed in Laos and Cambodia.
Technological and logistical abilities still reign supreme if it comes to an all out war. Asymmetric forces like with Aghanistan and Vietnam were problematic because of political pressures at home, and the long drawn out nature of those conflicts which increased political pressure. The 2003 invasion of Iraq was also perfectly executed, but the long duration occupation was the problem.
Asymmetric forces and insurgencies are problematic, but people overate them in modern discourse because of the weight of the occupation of Aghanistan and Iraq in recent memory. They are easy to crush, but politics at the time was complicated.
I don't know if I would say "easy to crush". For Western nations i think it's especially hard because the images of civilian deaths have a large effect on the population and voting. Fighting asymmetric forces seems to always involve significant civilian casualties. Other nations operating without those constraints would have an easier time destroying asymmetric forces IMO. I think they would use the same tactics used for crushing dissenting civilians only kill-on-sight instead of at least the illusion of arrest-on-sight.
/not a military/war expert
My point is that the oppressed population who "lost" (Iraqis, Afghanistanis etc in these examples) aren't just like "oh well - I guess I'll just abandon everything I know and believe in and start eating beef and alcohol since the US invasion wiped out my country's military.". Instead they are resisting and so now you get the quagmire where basically anyone old enough to stand is now a threat to the occupiers even if they "wiped out" the army.
So you're not going to have some hypothetical "robot war" where no humans are hurt. The humans will start getting hurt once all the robots are dead.
this is ancient warfare in a nutshell: a couple of battles where you clearly see who has the better army and the side which starts losing a bit too much surrenders the country
This assumes that human lives will be more valuable than war robots. However, humans may be used as cannon fodder to distract enemy robots from targeting valuable equipment.
https://en.wikipedia.org/wiki/The_Secret_War_of_Lisa_Simpson
But for now the humans are still better at actual soccer than a RoboCup MSL team :-) https://www.youtube.com/watch?v=pE4UopWe2lo
Robot-robot matches for some types of robot are getting more interesting though: https://www.youtube.com/watch?v=g9bLIztscyI
Something like volleyball or tennis which is explicitly a contactless sport? Sure, provided that robots use similar strength to push the ball, otherwise a lot of fingers will be broken. Football is never happening.
In that penalty example you've shared nothing happens and then the ball is in the goal. They're supposed to be much more physchological. Both players are supposed to pay close attention to body movement to "predict" the side and the goalkeeper is supposed to see the shooter approaching the ball to know when to react.
Human-on-human play is already injury prone.
However they'll win easily if they are made to be stronger and / or faster.
Strength: hard for humans to push them around and to steal the ball from and easy for them to do it to humans.
Fast: they'll just throw the ball past the defenders and run, like boys vs children. No need to use any soccer skill.
What's the intermediate step for this? What's the record for a bipedal 100m dash? For one big enough to meaningfully kick a soccer ball (even if it can't currently)? What's the 100m dash record for a robot that can also run 7+ miles over 90 minutes?
I feel like we're close -- but we're not really close. We're close on individual qualifications, but not combining them. We may genuinely be there with the control systems and computing if we had a performant robot.
edit: found the project homepage https://leggedrobotics.github.io/rl-blindloco/
That said, during competitions it is not allowed to use external motion tracking as it is used in the paper. The necessary image processing takes a good chunk out of the processing power budget and the quality of image sensors this size doesn't help. Also most teams wouldn't train their expensive robots this way because the risk of damaging it is way too high.
I think this is a glimpse of what's to come in robot soccer. But we are a long way away for self contained autonomous robots to behave this way.
They train it in a simulation, and these robots are quite expensive too. The only problem is getting miniaturized hardware, which we are bound to see very soon as there are vast practical applications, I'm thinking dishwashers and refrigerators.
The problem is imho that the problem is way harder and we will probably struggle to train a net that can do this in simulation. It might just be impossible with our current capabilities, we might have to wait a few years until research catches up with the problem. But I am not a RL guy and can comment only from my computer vision-perspective.
https://emanual.robotis.com/docs/en/platform/op3/introductio...
$14,199 on AliExpress.
We have seen interesting robots like this one, from Boston Dynamics and others but they still seem to be far from displacing humans in the workforce. It seems that these robots have similar fine motor skills to humans, enough to displace humans in more economic activities.
If visual detection and response is the problem, teleoperation is always an option: hire people from low-labor cost countries to operate them remotely.
So I am assuming that some kind of manufacturing costs are the reason that these robots are not taking over more human labor in advanced economies. Is that correct?
Can't claim lots of experience, but: Consider how much is happening with these kinds of robots that's not actually related to the task.
If you want a production line, a well designed arm with replaceable tool will be much more efficient. If you want a fast moving delivery robot, you want a shelf with wheels. If you want all terrain mobility, you have treads.
Currently it's much easier/cheaper to get something specialised for the task you actually want done rather than building something super generic and human like. "Can it balance itself while chasing a ball" is not a problem you want to be solving, ever. Neither "how to use tools while moving sensors can see the object from extremely limited angles". Human-like robots introduce lots of issues we don't need to be solving - why build a Rube Goldberg machine if you don't need one?
Every single joint/motor you don't need is extra maintenance. Every single "this thing needs to balance itself" element is wasted power. Every extra self-support processing from the sensor is unnecessary delay.
You could apply the same argument to computer hardware: General purpose CPUs are incredibly wasteful compared to specialized hardware (see Bitcoin mining where ASICs are many orders of magnitude faster than even GPUs), yet on the whole the value of being able to iterate in software wins out.
On the other hand, a generic humanoid has lots of cost/complexity/maintenance/etc impact. An articulated arm with a tool for a given task can do things you can't generalise (think DaVinci robot precision vs auto construction high power - humanoid can't replace either)
There's also massively higher market for CPUs (multiple per person, cheap) than for humanoid robots (tiny percentage per person, expensive), so the economy is different.
It's not manufacturing costs, it's that robots aren't very good.
Automated burger cooking doesn't put a burger on a grill/griddle and flip it when it looks good, it cooks it from both sides at once on a conveyor with human set timing. Sewing machines have a different method than hand sewing. Printers don't generally move a pen across paper, they make a series of dots in a line (old school plotters are an exception, but modern plotters are inkjet dots in a line). Dishwashers don't scrub dishes, they just shoot recurculated hot, soapy water everywhere.
Methods that would be tedious or difficult or unsafe for humans can be simple for machines, and following human methods would be difficult, imprecise, and costly.
A factory environment can usually be arranged with belts or tracks to avoid carrying, and with surfaces suitable for wheels rather than requiring legs.
6 axis are quite common, and getting every more common They were always used, but are now like, not the price of a house in the bay area for a reasonable one.
It's rare to need a lot beyond that for say, production line manufacturing for a wood processing plant.
Everything is already in line, etc.
They can lift and move things, spray finish, you name it.
Most of the time is programming.
The arms can be fast and accurate.
Now, the more general kind of warehouse work, like people walking around shelving things, that's harder - you get into issues around size of motors, actuation accuracy, etc.
In lots of cases it's easier to design a warehouse around robot capability than design robots around warehouse structure.
Unless there are dramatic breakthroughs, you will see arms and such get used more and more at places with less money and lower margins (IE random metal/wood shop) well before full on bipedal/quadruped robots.
Or you know, people will use the arms as novelties to smack the person who broke the build in the exact center of the back of their head with a precise amount of force.
Both seem equally likely ;)
Now that seems to be changing.
Then there are actuators. Small robots use servos which are not fast enough. Bigger robots, like those from Boston Dynamics, use brushless motors. Which are fast, but have limited torque.
Then there is the cost. Without volume it's absurdly high.
Then competitors, they can easily reproduce. For example Boston Dynamics sell it's dog sized robot for $75k, while mechanically similar Chinese are $4-10k.
Then the biggest question is what are you going to do with it for this price? Special note: we already have vacuums. You cannot send it shopping alone. Any other realistic ideas?
Basically, with current automated solutions humans form the "glue" between all existing home automation solutions, because humans are generally capable where existing solutions are specific and fragile. If that "glue" role can be performed by a humanoid robot, then it dramatically reduces the amount of housework humans have to do.
As for who might buy it for the price, the obvious answer is people who are rich enough to afford it but either not rich enough to hire service staff or who don't want to employ service staff for other reasons (privacy, don't want to be personally exposed to wealth inequality, racism, fear of theft, fear of disease). Another answer is disabled people. A lot of disabled people love their support workers and have important, humane relationships with them. A lot of other disabled people (like me) love what support workers do and think it's important but are unwilling to go through the arduous process of finding someone you trust to be there when you're at your most vulnerable, or being forced to make do with one of the small proportion of support workers who don't care about their clients that much or are borderline abusive (or actually abusive). In countries with socialised or subsidised medicine, I could very easily see the case for covering humanoid robotic helpers. Already in my country various machines worth up to $8000 or even $100,000 can be covered for disabled people depending on how much it would help, the medical need for it, etc, under various different programs. These aren't small programs either, a lot of disabled people here have access to these resources. A lot of money is also allocated on an ongoing basis specifically for support workers, and for humanoid robots specifically I could see them being eligible to access those funds because of the overlap in role between them and support workers. It won't be a good solution for every disabled person, but it could be a good solution for a lot of us.
Maybe I'd enjoy real soccer more as well, if the players just sometimes randomly face-planted onto the field!
I have no problem with robot football warfare settling disputes. It's the fan violence which worries me: Maybe we need robots for that too?
Basically its F1. The drivers are cosmetic at this point, and the fans are all behind TV screens. If somebody tells me F1 has been faked for the last 10 years I would believe it.
I like how they trained the robots to fake "AAAAAH MY ANKLE HE KILLED MY ANKLE REF REF FOUL" falls.
Ben Bova wrote a juvenalia SF about it in 1969 called "the Duelling Machine" -its classic coming-of-age stuff.
In times past, in history, yes you could settle things like kingly succession by a single pair of champions, but I find it passing strange. Its like the armies basically say: "stuff this, if you want this outcome so badly, you duke it out guys" -but this comes from a time when wars were won or lost by who could pay off the mercenaries the best.
"Lets take this to the international court of arbitration" only works so far. China (spratleys) and Australia and East Timor (oil wells) can tell tales about how much people behave nicely in international arbitration. They behave about as well as a robot trained to fake out ankle kicks HE HIT ME REF REF ITS A FOUL ITS A FOUL I WANT A FREE KICK
>In the distant future, mankind has forsaken global wars for battles of single combat. The world has been divided into two opposing super powers, with each side represented by trained champions.
Mechwarrior boxing, more or less. The 5/10 imdb rating is well deserved. Mostly notable for a writing credit by Joe Haldeman, who managed to get an accurate depiction of a megawatt-class laser weapon on screen.
Obviously it won't work every time. But we do have controlled settings to settle disputes.
Previously if two different persons thought they ought to control a country they gathered troops and fought it out. Nowadays in many modern countries we try to channel these disputes into "voting" and "democracy".
Similarly in days of yore if your boy disrespected mine we got a good blood feud on our hand. Nowadays we try to channel similar disputes into lawsuits.
The range of conflicts and disputes we settle with force vs the ones we settle with simulated violence is not fixed.
Why do people, very powerful ones tend to agree to use democracy or lawsuit instead of unrestrained violence? Because we arranged the circumstances such that to do otherwise would impart a greater risk and cost on them. And that is the same how we can make anyone honour the result of a "war in controlled setting".
It is indeed pricey though, IIRC the cost of a Boston Dynamics Spot-like robot came down to around $6-7k. There is still room to drive the price down though.
[1] https://jpieper.com/2020/11/09/mjbots-november-2020-update/
It would probably be $25k-$50k and 1-3 years of development time depending on your skill set.
Oh and another thing: the way they walk (and fall over) reminds me of the classic indie wrestling game Sumotori Dreams: http://www.gravitysensation.com/sumotori/
https://www.cbsnews.com/news/google-artificial-intelligence-...
We already have unmanned drones.
Creating expendable soldiers will be a huge win for whichever country achieves it first.
What is your plan for getting expendable soldiers for less than this? Even Russia spends more per soldier.
I can' wait to show my kids. This will be a "collision of worlds" to see what is possible, by bringing together their favorite sport (that they play constantly) and their lego interest... and what I hope is the long path of deep work, culminating in FRC .
Out of curiosity, is the field signaling to robots relative positions and goal locations via radio etc, or are the robots are working 100% with visual recognition ?
Then someone will get permission to let you pay for first-person or pseudo-third-person streams from given customized combatants.
This will be a definite "personalization of warfare" moment.
Then special ops of the kind only previously imagined will cross the traditional extreme privacy barrier.
A group of teens will crowdfund a tremendously capable soccer/spec-ops bot that strongly resembles Christ Jesus. They will dress it in robes, swim it into the Spratleys, and it will emerge from the water, tie a crown of thorns around its head like Rambo, and then...it will be mowed down.
But this will only be the first exciting moment. I fully expect version 3 to make it past the beachhead.
We just keep plugging along knowing that embodied control MDPs are the true path to AGI
:)
Safe and effective - that phrase gave me the gag reflex.
Or is this based on convolutional nets + RL or something else?
Also - a lot more flopping