Remember Atlas, the humanoid robot from Boston Dynamics? The company bought by Google, er, owned by Alphabet, and uh, most likely to become Skynet? Well — they’ve just shown us that Atlas can take a light jog through the woods now.
Published on YouTube a few days ago, Boston Dynamics gave a quick presentation on some of the upgrades the company has been working on for their bots. First up is a demonstration of Big Dog’s new appendage… What looks like an elephant trunk with a prosthetic hand on the end. Big Dog can now leave the testing lab any time he wants — door knobs are no longer an obstacle. Considering its been able to traverse extremely rough terrain for years now, this doesn’t bode well…
Atlas on the other hand has also come a long way. From standing on his own for the first time back in 2013, he moved quickly to being hit with medicine balls (and not falling over!) — and now, he can run outside. Luckily they haven’t quite figured out the battery pack solution yet… Video of his outdoor escapades after the break.
46 thoughts on “Taking Atlas For A Walk”
The uncanny valley of robotic gaits.
It’s doing that shit-in-pants walk because they -still- haven’t figured out an algorithm to do passive dynamic walking like real humans and animals do.
The robots are surprisingly simple and “dumb” when you realize what they’re actually doing. They’re doing very simple geometric things like planting two feet diagonally on the ground and then falling over the pivot, and the biggest development has been in accurate real-time measurement of the surroundings and processing that information fast enough to allow them to catch the robot before it falls over too far.
That’s why the big dog has to trot constantly even when it’s moving very slowly – it can’t just plant its legs down because the algorithm isn’t clever enough to figure how to move the body in any other way than tipping over a pivot.
I hadn’t heard “shit-in-pants walk” as a technical term before.
Seriously, nobody has figured out inverted pendulum gait yet? I figured that’s what all the inverse pendulum balancing robot projects were working toward.
They’ve done it in passive dynamic walkers that step downhill or on a slanted treadmill, but it’s apparently too difficult to calculate because it isn’t exactly an inverted pendulum – it’s a double pendulum going one way, but when the knee goes past a certain angle it locks up and the leg becomes a single inverted pendulum, making it somewhat chaotic and hard to model in a continuous equation. If the knee is locked, you can step on it, but if it isn’t then it will buckle under etc. etc.
So a robot like Atlas simply avoids ever locking the knee, and always operates in the mode where it can assume the leg does not suddenly transition into a stilt and back.
Part of the problem, I think, is that the robot would actually need some level of understanding of what it’s stepping onto. It can’t just plant a foot down and hope for the best.
For example, if you lead a human onto ice and they attempt to walk normally, they fall on their asses. When they know to except the ice, they too take a slightly bent-knee approach to it to maintain a tighter base of balance to avoid lateral forces under the foot.
The robot is essentially walking blind on ice all the time because it’s AI is not intelligent enough to make predictions about what will happen at each step. The fact that it stays upright at all is because it’s incredibly fast to react and correct.
It follows the general trend of AI research: if you have no idea how it’s done, just do something really fast until it looks like intentional and controlled and smart.
To support the notion:
“Left uninstructed, blind children develop awkward walking patterns that draw attention to their unusual gait. This “blind gait” is characterized by a flat footed, shuffling pattern. Each step is small, the feet swing outward, and the base of support is wide. The effect is almost a comic waddle.”
Remind you of something?
Don’t worry about the crappy gait, even if our best and brightest fail to find a good algorithm, the military has got our back:
Anand and his colleagues claim to have reproduced 99% of the brain’s diverse cell types and genes. They say their brain also contains a spinal cord, signalling circuitry and even a retina.
The ethical concerns were non-existent, said Anand. “We don’t have any sensory stimuli entering the brain. This brain is not thinking in any way.”
Good to know that according to the military sensory deprivation means not thinking. That some guantanamo captives seem to remember themselves thinking, and what they were experiencing during sensory deprivation hence must be seen as a lie. One must not see it as a counterexample of this definition of consciousness.
So in the future patriotic soldiers will only have to donate some skin cells, which will be turned into stem cells, from which brains will be grown (including spinal cord & retina). Retinas are already connected to cameras to make blind people see again. If you have to choose between fighting or shedding a few skin cells what would you do? Who cares about your asexually reproduced brain child?
The brain-robots will not suffer from having culture, ethics, and other non-military nonsense. Their virtual “youth” can be precisely controlled, and multiple brains with identical genetic makeup will allow to factor out genetics from environment, so that influence of specific forms of training can be objectively compared (i.e. identical youths except for a specific difference in training). Of all the donated skin cells, the best and most obedient can be selected, or perhaps one day converted into an immortal cell line.
It is also a win for the moralfags, because in the meantime the military can pretend to be solving Parkinzheimer or such and so.
So let an international market develop the hardware, but be the first to grow brains and efficiently use the hardware, good thinking.
Also, biological brains have this innate concept of parents to blindly trust, obey and mimic, and are optimized to learn from them. Artificial neural networks do not have this trait which was acquired over many generations…
If you have mastered QWOP, then you have figured it out.
The only place I’ve heard “shit-in-pants walk” was in an episode of ‘The Venture Brothers’. https://youtu.be/Scx-uM7mIfE
I,for one, welcome our new pants shitting overloads.
I doubt it’s an algorithmic issue. Compared to humans, perhaps it lacks some of the degrees of motion. Those it has are not as limber, fast, or well-controlled. The algorithm is adapted to the physical constraints.
I don’t know much about their design, but most robots lack a sense of weight/strain. Humans and animals can feel how much force they are applying which gives them immediate feedback on weight distribution. That, combined with other balance sensing mechanisms, makes it very easy for humans to stay upright and walk. Not to mention, human walking is not algorithmic, but learned.
Human walking as I understand it, comes pre-programmed and is inherited. The way you walk is almost like a fingerprint, and can be used to identify people.
There’s an inherited component that is literally in your spine – it’s automatic – which is modified and overridden as necessary by the brain. In prosthetic legs for example, you can go a long way by simply repeating a canned motion sequence in synchrony with the person’s healthy leg.
I certainly agree a lot of characteristics are inherited, but a lot is also learned.
You yourself gave the example of blind children, if the brain control of walking was so inherited, wouldn’t these blind children walk more normally just as their non-blind parents? This triggers my bullsh*t alert.
I believe the walking of healthy humans is inherited (but only indirectly so), after all animal species (each with their own genome) have specific walking styles. I think the causal path is more genetics => skeleton,muscle,weight size and distribution => brain adapts to find most efficient way to walk. So from the perspective of the brain, it is just solving a general problem. After all people drive vehicles (bicycles, cars, skateboards) both in the real world and in games, and easily train themselves to drive them in a way the vehicle feels part of their body. (even populations that were just a generation ago too poor to drive cars, learn to drive them without problem today, … or would you claim their genome is clairvoyant, and they inherited what their forefather genomes predicted to come as extensions of the body? or that natural selection somehow works within the lifetime of the individual -population size 1- without needing genetic diversity? like the old theories that “when trying to reach higher parts of trees to eat from giraffe necks elongated, proportional to the effort the giraffe puts into it”? what a crockload)
It just needs a hydrogen fuel cell battery. May be 2 of them for redundancy. (c:
Does it really has higher power density (in relation to mass) than combustion engine+generator that was used in BigDog?
i think he was just making a Terminator ™ reference
That’s right. I don’t care how good its synthetic skin is; it’s not going to infiltrate many human rebel bases, if it sounds like a lawnmower with an exhaust pipe coming out of its bottom. (c:
you should meet my mother-in-law.
Just run’em on Alcohol like Futurama!
How do you refuel in the field? What about the heat signature? Bender (Futurama) drank ETHANOL, not sure it was for power supply but more of a rust inhibitor. :-)
The Future is Stupid(tm)
Honda’s Asimo walks the same way “S.I.P.W.” like DAX above said.
Why did SONY Kill Q.R.I.O.? 120-degrees of freedom movement…
I suspect because it was more animatronics rather than useful research. In some ‘blooper’ videos you can see some fall over and continue the pre-canned motions on the ground like toy robots.
Have you ever seen the Sony QRIO interact with children? It has a voice synth and recog unit, it’s cpu is radio-linked to an AI computer, it has fluid movement with it’s hands and feet like it could wield a small hidden firearm deftly. It could be rigged with a rut-sack explosive charge, etc. When it falls it gets up with ease. These things do NOT need a human handler at all – just a long range strategic operational plan. They can venture out into the field disguised as cloaked children (with a hoodie or something). They can tactically interact verbally with indigenous informants in their language. They can be equipped with on-board GPS to follow terrain to target. They are more like a cruise missile only with feet. The head looks deceptively friendly. I wonder who might be considering this package for paramilitary deployment?
Operation Rut-sack [sic] in the Cold War period in East Germany used American military people to carry very deadly backpacks.. QRIO could eliminate the risk and penetrate deeply into enemy territory as a seemingly innocent child and set off it’s rut-sack charge. Or collect intelligence and phone home much like the StarWars Imperial Viper Probe:
Hydrogen Fuel Cells are problematic. The Germans are using them for their new stealth submarine fleet. The problem is just like the Viper Probe video above shows at the end. It must purge fresh water to reduce weight from the collection tank. HFC’s expel water as a by-product. That can be tracked by the enemy. QRIO uses a phenomenal battery of some sort. Maybe some sort of lithium combination. It could recharge via sunlight (a kid sunning itself?). Atlas would need a very quiet fuel source that does not reveal it’s location like spore when on-mission. Solar-recharged batteries seem to be the only thing that fits the bill. They leave no traces of their existence. And Atlas is a huge target in the field. Really shows up on FLIR and RADAR too (too much metal). QRIO has little metal, mostly PVC or nylon joints.
Oh another power source is the RTG. However, it is highly radioactive. But it’s small and quiet. However, it gives off a hell of a lot of heat signature for FLIR units to pick up:
Somebody in SAN DIEGO CA is looking at QRIO (et al) for it’s ability to fit the bill of a NUGV (Novel Unmanned Ground Vehicle). The mobility envelope of a robot or NGUV (like ATLAS) within which they might expect American soldiers to perform will surely include some of the following:
· Traversals on a planar surface at rates less than 4 minutes per statute mile.
· Balanced traversals for 1hundred feet across a horizontal 4-inch beam.
· Vertical jumps over a 7-foot bar.
· Vertical climbs of a 30-foot rope
· Vertical climbs of a 50-foot ladder or wall with foot and hand holds.
· Horizontal jumps over a 28-foot span.
· Crawls beneath or slips between a 10-inch space.
· Swims in sea-state-1 for 1 mile.
No robot or human can do any of this YET…
Any robot from boston dynamics would stomp those sony toys into dust though.
And I doubt they could even carry an orange, unless it’s made of foam.
It’s not how much it can carry but how it moves quietly, swiftly, and the lightweight payload it could carry to target. It could effectively carry a handgun designed by Heizer Defense (i.s. DoubleTap only ounces in weight with 4-rounds) and/or a few ounces of Corpent (lightweight powerful plastic explosive) in it’s infrastructure. But you wouldn’t have to use QRIO as a robot assassin like how General Atomics MQ1 robots are used today in SW Asia. No you could use QRIO as a spy-bot like the Viper Probe I posted above. Boston Dynamics only makes rugged rescue bots, there are not many sociopathic MIL-IND-COMPLEX Defense Industries like General Atomics, Northrop Grumman, and others. Stephen Hawking and Apple co-founder Steve Wozniak, have issued a stern warning against the development of so-called killer-robots.
That hasn’t stopped Northrop Grumman from developing this monster MADSS (CaMEL):
I think Northrop Grumman and the like have been developing these things for ages, long before that initiative to prevent killer bots.
In fact such initiatives are the outcome of seeing all that stuff really.
Sorry about the dead link above. Also QRIO on has 27-degrees not 120 like I said above.
Umm, I think I might have a problem with one of the interpretations of the phrase “…we are having pretty good progress on making it so it has a mobility [that is] sort of within a shooting range of yours.” (1:01-1:08) :-P
PETMAN and Samsung’s Roboray and some others have been doing straight knee walking for some years now. ATLAS is behind the leading edge there. The full text of this research paper is available for free. http://www.researchgate.net/publication/261390206_Human-like_walking_with_straightened_knees_toe-off_and_heel-strike_for_the_humanoid_robot_iCub
The straight leg gait doesn’t need as large/powerful knee and hip motors so the legs can be smaller, which further reduces the leg power requirements.
What these robots lack, which would drastically improve their mobility, especially when turning, is the ability to twist their legs below the knee like humans. Without that, to turn they have to shuffle around by twisting their hip joints.
In my younger years I used to run in the dark across uneven ground, without falling. ‘Course I’d only do it where I’d previously made sure there was nothing up high enough to trip over and no holes to drop a foot in and catch it, bit still nowhere near as flat as a floor.
How did I manage it? By allowing my ankles to be loose and using my feet as angle sensors. I’d strike first with my heels then let my feet slap down with loose ankles until they were firmly planted. Then I’d tense up as if I was running normally and pivot forward over the ankle and spring off my toes. Initially it took some concentration but soon became easy, especially after running a path a couple of times. It’s a trick that without 3-axis ankle joints (or twist somewhere between knee and ankle), powerful toe spring-off capability and a straight leg gait would be impossible for a bipedal robot.
Another running trick I used to do, I called ditch running. We had a large hill with a flood irrigation ditch along the top. The ditch was about a foot to two feet wide and six to nine inches deep. I could run along that ditch, planting my feet on the opposing slopes of it, criss-crossing my legs back and forth. I never got my feet tangled up doing that.
I could also stand for a long time on one foot, either one. To me it was simple, just pick up one foot and tweak muscle tensions at hips and planted foot/ankle just as much as required to maintain balance. Meanwhile the rest of the P.E. class would be flailing about, having to put their lifted foot down, or even falling down – and looking at me, wondering how I could just stay there standing still on one foot with my arms at my sides instead of waving them about.
Trying to keep the vertical balance via ‘external input’ of waving the arms and free leg around is akin to Pilot Induced Oscillation. The more you try to gain control, the more you over control and the faster you land on your backside.
It’s all simple, for a human, if you pay attention to how your joints and muscles work.
My idea for a starting point for a humanoid robot would be to start with a precision replica of a human skeleton. Study that to see how the joints move. The hips are the simplest, they are free rotating within their limits around a single fixed point. Shoulders are that simple at their pivot but that pivot is held in place in a shallow recess on the scapula by some cartilage and tendons. The clavicle is fairly securely attached to the clavicle by tendons and held down to a small recess at the top of the sternum. The scapula is just held in place by muscles and tendons. So a human shoulder has a lot of degrees of freedom, at the cost of being rather “squishy” and imprecise in location – thus easily damaged. That’s why the clavicle, being the most securely attached piece of a flexible system, and paradoxically the smallest/weakest piece, gets broken when you fall hard on a shoulder. It’s somewhat like the front suspension on a car, if the upper A arm only had the front leg and the lower A arm only had its back leg with its mounting bolts left loose, with its outer end attached to the top of the spindle and the lower end of the spindle flapping loose – with a lot of springs and rubber bands attempting to hold it all in place to allow the car to be driven down the road.
The other joints have one or two degrees of motion that are not around fixed points. Their rotation radius changes as the joint moves. Robots with knees, ankles, elbows and finger joints that have fixed rotation points look odd to us because their motions, even when very fluidly programmed just don’t look right. That non-point rotation caused problems with artificial joints when early designs used fixed axis hinges. The tendons and muscles and other tissue isn’t made to work that way. Then the process of relining knee joints with plastic and metal shaped like natural bone ends was invented and they work much better. More recently that design has been used for finger joints, cut off damaged or diseased ends of the bones and replace them with artificial parts shaped like normal bone ends.
I think PETMAN was more non-mobile and used for chemical and biological-warfare testing. Not sure. I could be wrong. That S.I.Y.P. walk that DAX refers to is necessary to maintain balance. I know it looks goofy but senior-citizens do it all the time when they walk. It does help you stay upright. Of course the cane helps too:
Read the paper: it offers the explaination that a straight knee and flat foot step causes a singularity in the system that basically prevents the robot from taking the next step.
So basically, the Atlas robot keeps the hip at a constant height and shuffles the foot underneath as if you had a cart on wheels as simulated by the legs, whereas real people bob up and down, essentially pole-vaulting over the front leg with every step with the help of the heel-toe and hip-roll action which is lacking in the robot.
Can’t help myself, but the DARPA robotics projects always give me the creeps. Where most Japanese projects aim towards making daily life easier with an extra humana touch, Atlas is definitely more on the gunslinging T800 side for a “rescue robot”.
Please remember what the ‘D’ in DARPA stands for. Atlas will not be used in OFFENSIVE operations. We have “other” agencies who seem to have sociopath-like agendas much like what you propose. Try and remember who concocted the once top-secret L.R.V. program of the 1960’s. Read what their offensive agenda was in the November 2000 Popular Mechanics article:
Was just thinking that a Sony QRIO could be disguised as a indigenous desert child like the JAWAS in the Star Wars movie. They could be just as deadly as the Jawas were (LOL):
Here is Sony QRIO communicating with children with it’s Autonomous Behavior Control Architecture. No human controllers or handlers. Imagine putting JAWAS hoodies on these things. The incognito war fighter on LRRP could carry one like a child in a papoose. Then deploy it at key positions when needed. That way it won’t look out of place without an adult guardian nearby. Also save on battery power trying to cross a desert on foot. It would never need food or water just sunlight. It could communicate with indigenous children like asking: “Hello! My name is Abdhul. I’m looking for my uncle Anwar. Can you tell me where he lives in this village?”
If QRIO’s JAWAS-like glowing eyes and goofy looking head is a problem for interacting with indigenous children during an in-country LRRP mission, then replace the head with a DIEGO-SAN head. University of California at San Diego is working on an autonomous robot that displays realistic child-like facial emotions. Imagine this thing interviewing a village child for locating a hard target. Diego-San will be walking the streets of San Diego soon (kind of creepy):
Currently in most warzones being a child really doesn’t; offer much of a protection from harm though.
and if it did, that would be over in 2 days after someone deployed a child-looking killer robot.
And now I”m reminded of the scifi movie ‘screamers’. If you didn’t see it yet; go get it..Based on a story from Philip K. Dick and staring peter weller.
Since QRIO would be equipped with a 23 oz side arm with 4 .45 caliber rounds. a parent-soldier nearby (in disguise too), and a few ounces of plastique in its gut, it wouldn’t need much protection from insurgents. Couple that with a Kevlar hoodie and his outer hull is the new Moratex STS better than Kevlar. It can also jump, run, and hide. In 2-days who would be left alive to tell of a child-robot assassin? Once he enters the hut and picks out facial recognition targets, says “Are you Anwar?” He says “Yes I am child…” then kaboom. Everybody would think the hut was hit by a Predator Drone strike and not a tiny killer-robot. The explosion would take out about a quarter block radius.
But why use them for killing? Why not a clever spy-drone with all sorts of sensors? He could just send his telemetry back to his soldier-escort and not a satcom link. The interviewing skills would be controlled by the soldier and he or she could hear the entire conversation. Of course it would be in Arabic, Farsi, Urdu, or Pashtun,
Yes I was trying to remember a movie like this. Screamers huh?
I just want to see an ATLAS in Freddy Fazzbear cosplay before the FNAF series is totally forgotten.
Freddie is a animatronic stuck on a pedestal. ATLAS can actually walk and run, He also has no scary face like Freddie. They haven’t thought of one yet.
Please be kind and respectful to help make the comments section excellent. (Comment Policy)