Boston Dynamics spends months training its Atlas robots to perform one minute of parkour almost perfectly Boston Dynamics has released its flashiest promo video yet: a one-minute clip of not one but two Atlas robots competently completing a parkour obstacle course. At first glance, it’s simultaneously impressive and bizarre. The pair of fridge-sized machines move fluidly, jumping over blocks, stepping up platforms, and walking …
A significant difference is in the memory functions of the two entities (human and machine).
Explicit Memory vs Implicit Memory
Humans have both types of memory, and use their unconscious memories (implicit = non-recall).
The bots only have the explicit memories.
So the meatbag has the advantage, ergo the shortening in the learning curve human vs bot that you mention.
[There other factors that give the human the advantage..... for now muahahahaaaa!]
I have to wonder how much paralleliism there is in the robots' learning algorithm.
With a lot of parallel operations, you could power it with a server farm and something a bit better than wifi networking. That'd speed the learning process up, right?
(eventually >256 core CPUs but who knows when THAT will happen... at least one that STILL fits in a robot)
That's not parkour! I'm 55 but I could easily do everything in that video except the backflips at the end. I'll bet I'll be able to do it when I'm 75. That stuff was hardly more difficult than walking on stairs.
As for real parkour, I'll bet I could learn it in a month (when I was younger) if when I fall and break my leg I could have a new leg attached and be back at it 20 minutes later. The reason it takes humans longer than a month to master is because if they get hurt they may need weeks or even months of healing before they can get back at it.
If you were freed of any concern for pain or injury, you could learn parkour quite quickly. Given that it took a month to teach these robots to do nothing most of us can't do other than the backflip, how long would it take for them to learn REAL parkour? How much longer to do it dynamically rather than repeating the same pattern?
We certainly don't need to fear being chased on rooftops by these "terminators". They'd be lost if faced with a roof they hadn't spent a month training on!
Yeah, you forget they're robots and start complaining about how heavy footed they are. "They're like a tribe of dancing elephants, only one one-hundred-and-twenty-fifth as cute."
I was going to make a joke about them being ATLAS and P-body. But they damn well called their robot Atlas.
"it's still a non-intelligent machine that does one thing"
The whole point is just the opposite. They are trying to develop something that can do more than one thing. Theoretically at least, it has been told some basics and instructed to perform this routine which it has subsequently 'learned' how to do for itself.
Next they'll have it realise it's getting low on power, and use an extension lead to plug it's groinal attachment port into a 13A socket ... then do the hoovering and ironing.
you'll have my money. Until then, Meh.
(Only joking. This is immensely impressive. With a tonne more ML, it will manage the same objects in a natural environment. I wonder what will come first: completely natural object recognition using ML, or LIDAR based object recognition, or both).
I've noted it was a joke BTW, but
House work would be too difficult as an early task, homes are varied with too many variables.
I don't like my socks in a ball for instance after washing, and my sink tap sometimes needs a gentle whack and yes i like walking around that "obstacle"
Much better in a more structured environment like a Amazon factory removing lots of (expensive) humans jobs.
Or Cleaning the streets (day in day out the same streets, walk through once and repeat!)
No need for them on checkouts since we are becoming more used to doing the checkout operators jobs ourselves , they could replace the Herder who tirelessly directs us at the supermarket away from the single checkout operator still working to the bank of "Self Service Checkouts" with a simple Robot though but would we listen to a mere machine.
I would listen to a "Herder" robot if it was backed up by a Human OR it was ED-209 from RoboCop
For us it's natural but being able to use two legs and stay balanced while moving and using force is incredibly difficult. At the same time, it is an extremely practical ability when working outside factories and office space. Imagine how useful this ability would be, for example, for exploring foreign planets, though no doubt the military will have first pick.
And the research project and presumably the relevant patents will be around the necessary components and AI models that are required for this.
For the military, they'll be the ultimate drones - highly capable remote-guided weapons platforms with just enough autonomy to get from A to B while the operator maintains situational awareness. It's the ultimate in boots on the ground without risk to life, and even better if one operator can control the coordinated actions of an entire platoon.
For exploration, it's the ideal telepresence platform - cocoon your humans in a safe orbital station around Titan or a ground habitat on Mars and send the bots into the hazardous environment.
Until they're batteries run out of they get caught in a simple trap, stripped down and sold for parts in the local market.
Of course it could be a cunning plan. Make your military tech so complex that when you lose and have to 'strategically withdraw' leaving all the kit behind, the locals can't reuse any of it.
Better hope that Toyota doesn't start making robot pickups
The military was indeed very impressed by early models but the energy problems: batteries don't last long, petrol engines have their own problems quickly put paid to them as soldiers. But for reconnaissance or camp work it's possibly a different story.
Boston Dynamics has been "parodied" and are definitely fakes, but the real company has released valid ones too:
"Bosstown Dynamics": https://www.snopes.com/fact-check/armed-robot-video/
Boston Dynamics: https://www.snopes.com/fact-check/real-dancing-robots/
Still, the future is coming fast, both in robotics and deepfakes.
Probably longer than that; maybe twenty minutes of full performance, but with reasonable power management, possibly as long as an hour?
And of course there's no information on battery life publicly available- that's a trade secret.
(my guesstimate is based largely on power usage from how fast some of my cordless tools can drain the larger capacity battery packs I have for them.)
You make it sound like having the motor skills of a toddler isn't impressive - and I've not seen many toddlers who can back flip or jump that confidently and accurately.
Beyond humans there are very few animal species that could do all those actions on just two legs without a tail.
In fact whilst most adults and older children could do the steps, sloped wedges and transition from the blocks to the beam that would be done with more of a stepping jump with a leading leg. I doubt many could do the transition as three, two footed standing jumps without any swaying, repositioning or needing to regain balance on landing.
Don't knock toddler's motor skills. From a college textbook on child development, I read about an experiment where adult volunteers in peak physical condition (I believe they were recruited from an armed forces unit) were tasked with copying all the physical movements of toddlers at play. The grown ups lasted something like 20 minutes before they were exhausted.
(Granted, there's stuff like the square cube law at play, plus a distinction between motor skills and stamina)
In part covered by my post above ("explicit memory vs implicit memory")
I'll add to that that humans are here as, so far, we haven't failed a Darwin test. The bots we're making have to rely on us to be their "natural selection". We'll (they'll?) get there - if it doesn't work, try something else. But the evolutionary curves are at a very, very different point on said curve.
Yeah, it'll do the housework... loading the kaftans into the washing machine, cleaning the teacups, sharpening the beheading scythe, picking up the stones for that evening's stoning - that sort of thing. And after just a few minutes of training and an AK47, it can mow down any 'non-innocent' locals who happen to be passing.
Battery life? That's important to know.
“That is to say, as far as we can tell, the robots were given a set of basic actions by their creators, and then learned how to use those actions to get from A to B from what they could see around them.”
The same search for the right solutions in this context: there is a context and a solution is searched for. There the context should be annotated in a proper manner, which makes it searchable. This is the AI tecnology.
Whilst the parkour demo is impressive, a more useful form factor than humanoid is the centaur, and given the already developed "Big Dog", it shouldn’t be difficult to build. The added body could accommodate additional power sources, support further cargo weight, and might allow for horse-like running speeds. Such a robot form could be the basis for soldier 'bots.
By completing a parkour obstacle course, two Atlas robots from Boston Dynamics, demonstrated the power these machines can play in the future.
These wonderful machines jump, walk, and climb over platforms, blocks, and wooden planks, showing us a frenzy of multiple actions.
The achievements of these machines are the hard work of engineers and scientists.
The movement from points A - B, from physical environment is not mere virtual theory, but practical theory.
Whether Atlas robots will be the future, nobody knows yet.
But the future may be just around the corner.
Reference
Katyanna Quach. Boston Dynamics spends months training its Atlas robots to perform one minute of parkour almost perfectly. https://www.theregister.com/2021/08/18/boston_dynamics_parkour/
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
