human-like gaits mainly require the correct structural elements (rather than machine learning), like locking knees (as noted in the article) and slightly elastic legs and joints (for energy storage/release as well shock absorbtion).
most robots are energetically inefficient and require extraneous computational power because they don't take advantage of passive dynamics (https://en.wikipedia.org/wiki/Passive_dynamics). it's disappointing to realize that we still haven't caught up to fundamental biological systems yet, even though the research is 2+ decades old.
"Preflexes are the latent capacities in the musculoskeletal system that auto-stabilize movements through the use of the nonlinear visco-elastic properties of muscles when they contract.[1][2] The term "preflex" for such a zero-delay, intrinsic feedback loop was coined by Loeb. [3] Unlike stabilization methods using neurons such as reflexes and higher brain control, it happens with minimal time delay. Its chief disadvantage is that it works only to stabilize the main movements of the musculoskeletal system."
somewhat related to that, our tendons are tuned to our gait dynamics to return the optimal energy back to us as we locomote (more accurately, our strides optimize around tendon elasticity and leg length).
Exactly. The thing that hasn't been figured out yet is how to combine the energy efficiency and aesthetics of passive walking with the versatility needed to have the robot do useful work, especially in a principled way.
in general, yes (natural selection is amazing), but the knowledge in this case has been readily available for literally decades but not applied appropriately.
organisms have evolved over billions of years to take advantage of passive dynamics, which is basically the application of newtonian physics to what is mostly a mechanical design problem. we've known for a long time that passive dynamic elements are not only energetically more efficient, but also makes computations simpler.
by contrast, active dynamic locomotion is akin to putting a joint between the front fork of a bike and the shaft, and installing a hand pump that you have to constantly actuate to keep the bike stable while you're riding it. you could do it, but it's not the best solution.
The trick is in the application. Most of our "industrial lego" is for making mostly rigid things with some spinny parts and a few, mostly passive sproingy/squishy bits. However, we ourselves are mostly sproingy/squishy and a lot of that is dynamic.
By the same token, most of the tooling and parts for building airplanes isn't all that useful for building orinthopters.
So it's not just knowledge. It's knowledge expressed in the form of industrial infrastructure and products.
> but the knowledge in this case has been readily available for literally decades but not applied appropriately.
This isn’t a kinematics homework where you can draw the links and joints and get full marks. You have to actuate them. For a human gait, you need to generate the same torque, at the same velocities, and the mass of the actuator needs to distribute across the linkages in the same proportion. Guess what? Actuators like that are hard to find.
This isn’t a kinematics homework where you can draw the links and joints and get full marks. You have to actuate them.
I should paraphrase that for interviews. "This isn't a CS homework where you can cite Big-O and get full marks. We have to be able to implement code that uses them."
And by that, I don't mean that you have to write all of the actual code. But one needs to be able to specify it exactly so someone could implement it and know enough about that to tell someone of the pitfalls in the current problem context/framework.
the proper kinematics makes both the mechanical design and the control systems simpler, including leeway in choosing the actuators. you’re right in that we don’t have muscle-like fast-acting axial actuators with pliable multi-insertion endpoints, but you don’t need to exactly mimic a human leg (which is effectively impossible right now) to apply these principles.
Whenever I see an Atlas video it surprises me how much I anthropomorphise it. I should be thinking about it like a wheel or jet engine or something, as it's just a means of locomotion, but I can't help by thinking about it like an almost-sentient AI, even though it has nothing like intelligence.
If you showed me a chat bot and a walking robot, I would say that the walking robot is orders of magnitudes closer to being human, even though it's just a glorified Segway.
This is similar to what I experienced. There seems to be a heuristic in the brain that sees a human-like object and immediately starts trying to superimpose other human-like qualities on it. It's almost like an optical illusion where the brain totally misinterprets a visual cue, but in this case it's perhaps more like a social/empathetic/mirroring illusion.
It is very interesting. I think our brains are just so incredibly fitted towards interpreting entities whose shape and movement is life-like as intelligent, purposeful agents. Even-moreso if they are human-like. You could give just-so story that this was very important for survival in our evolutionary history, which is a fairly compelling idea.
"Even Boston Dynamics’ own Atlas uses this crouching sort of squat-walk to get around, because those perpetually bent legs are how it keeps from falling over."
If I ever lock my joints I feel like I'm going to fall over too. I wonder if a more human walk might just be less efficient for bipedal movement and balance.
Also, I'm betting that by March we will see Atlas do a Fortnite dance.
For walking, I imagine you do lock your knee as you stride off of your toe.
That's a dynamic situation though, locking joints in a static situation pretty much requires thinking about balance, so you'd notice perturbations more.
Every time the grad student punches the robot with a stick I wonder when it is going to punch back :-) It says something about the improvement in anthropomorphism I think.
Real gaits though really need a foot with four degrees of freedom rather than the more common three. Typically the 'foot' plate has a roll, pitch, and yaw component. If you add an additional bend access where humans and animals have the 'ball' of their foot so that you have a flat surface on the ground while "up on your toes" to can add some additional gait styles.
Another reminder of how incredibly powerful our brains are. Walking and motor skills seem like such basic things because we start mastering them as toddlers. Meanwhile, teams of experts across the world are still trying to figure out how to get a robot to do it. For example, they're talking about finding out the location of where to toe-off based on some instantaneous capture point. None of that ever consciously came to my mind, but these crazy calculations are constantly running in the background.
I think it is a bit generous to call an AI's behavior in those areas "high-level reasoning" :)
It is certainly quite impressive. But it seems to me that the paradox is neatly solved by comparing the how well-constrained the activity is.
Movement and (especially) visual processing are incredibly general problems. Where as board games are very narrow, well-constrained problems.
It always interesting when people expect AI to robustly and effectively process visual content without first actually understanding the world (which obviously no AI is even close to doing).
"We’re trying now to design approaches that are capable of both precise footstep placement, such as when walking over a rock field with few, sparse footholds, and are robust to when this precision fails, such as really compliant terrain with lots of subtle height variations, using a single algorithm."
Somehow I have the feeling that robots capable of striding confidently over mountains of human skulls could be coming sooner than some think...
However, if the government would like to ramp up the level of military robot indignity, I guess we'll need some kind of ministry to track and catalog silly walks.
most robots are energetically inefficient and require extraneous computational power because they don't take advantage of passive dynamics (https://en.wikipedia.org/wiki/Passive_dynamics). it's disappointing to realize that we still haven't caught up to fundamental biological systems yet, even though the research is 2+ decades old.
"Preflexes are the latent capacities in the musculoskeletal system that auto-stabilize movements through the use of the nonlinear visco-elastic properties of muscles when they contract.[1][2] The term "preflex" for such a zero-delay, intrinsic feedback loop was coined by Loeb. [3] Unlike stabilization methods using neurons such as reflexes and higher brain control, it happens with minimal time delay. Its chief disadvantage is that it works only to stabilize the main movements of the musculoskeletal system."
somewhat related to that, our tendons are tuned to our gait dynamics to return the optimal energy back to us as we locomote (more accurately, our strides optimize around tendon elasticity and leg length).
Given that we're the result of billions of years of natural selection, shouldn't we expect that?
organisms have evolved over billions of years to take advantage of passive dynamics, which is basically the application of newtonian physics to what is mostly a mechanical design problem. we've known for a long time that passive dynamic elements are not only energetically more efficient, but also makes computations simpler.
by contrast, active dynamic locomotion is akin to putting a joint between the front fork of a bike and the shaft, and installing a hand pump that you have to constantly actuate to keep the bike stable while you're riding it. you could do it, but it's not the best solution.
By the same token, most of the tooling and parts for building airplanes isn't all that useful for building orinthopters.
So it's not just knowledge. It's knowledge expressed in the form of industrial infrastructure and products.
This isn’t a kinematics homework where you can draw the links and joints and get full marks. You have to actuate them. For a human gait, you need to generate the same torque, at the same velocities, and the mass of the actuator needs to distribute across the linkages in the same proportion. Guess what? Actuators like that are hard to find.
I should paraphrase that for interviews. "This isn't a CS homework where you can cite Big-O and get full marks. We have to be able to implement code that uses them."
And by that, I don't mean that you have to write all of the actual code. But one needs to be able to specify it exactly so someone could implement it and know enough about that to tell someone of the pitfalls in the current problem context/framework.
If you showed me a chat bot and a walking robot, I would say that the walking robot is orders of magnitudes closer to being human, even though it's just a glorified Segway.
Very funny how that works.
I figure that’s probably what I’m imagining when I see theee videos.
If I ever lock my joints I feel like I'm going to fall over too. I wonder if a more human walk might just be less efficient for bipedal movement and balance.
Also, I'm betting that by March we will see Atlas do a Fortnite dance.
That's a dynamic situation though, locking joints in a static situation pretty much requires thinking about balance, so you'd notice perturbations more.
Real gaits though really need a foot with four degrees of freedom rather than the more common three. Typically the 'foot' plate has a roll, pitch, and yaw component. If you add an additional bend access where humans and animals have the 'ball' of their foot so that you have a flat surface on the ground while "up on your toes" to can add some additional gait styles.
[0]https://medium.com/@froger_mcs/moravecs-paradox-c79bf638103f
It is certainly quite impressive. But it seems to me that the paradox is neatly solved by comparing the how well-constrained the activity is.
Movement and (especially) visual processing are incredibly general problems. Where as board games are very narrow, well-constrained problems.
It always interesting when people expect AI to robustly and effectively process visual content without first actually understanding the world (which obviously no AI is even close to doing).
Somehow I have the feeling that robots capable of striding confidently over mountains of human skulls could be coming sooner than some think...