Embodied Intelligence – Stanford Robotics Seminar Winter 2025

[00:10] about embodied intelligence. of over the past few months-ish.

[00:24] We recently started to think about the possibility basically to think about embodied intelligence

[00:39] So embodied intelligence, I'm sure you guys have heard Perhaps it means many different things. So from the context of this talk,

[00:56] we are pretty much talking about some level of programmability or decision-making through hardware system, and tunable materials and structures.

[01:12] useful from different perspectives. You could argue this kind of purely mechanical systems that can do computation or some sort of actuation

[01:27] So the idea is there are certain important lock you can only unlock when you are physically being present. And actuators are useful for robotics, especially

[01:44] electronic-free robotics, for extreme conditions And we in our lab also try to think be basically leveraged in a sustainable application.

[02:01] be able to harvest the ambient energy fluctuation from the environment, and maybe eventually degrade into the field.

[02:16] So with this in mind, I stitched together a presentation to talk through again some of our team's recent reflection in terms of what are the unique sort of applications

[02:32] intelligence, again through the lens of morphing materials So there are four aspects I want to touch upon

[02:44] with a bunch of examples that are from our lab. So embodied intelligence in the context We design a lot of morphing materials.

[02:58] as material systems that carry hidden forces that are usually calculated or predicted based on some computational algorithm.

[03:10] physics plus algorithm. mechanism, and you think through how do inverse design by leveraging these kind

[03:26] So through the years, we've been looking into different material systems that could leverage this mindset for design and engineering all the way from shape-memory plastic

[03:40] to pneumatic elastomeric system to even clay and even edible And most of the projects share the commonality of physics

[03:52] So here are some quick examples. So for example, we once looked into basically the idea of a residual stress release inside thermoplastic

[04:07] can be seen as a way to engineer self-folding and self-assemble So essentially, we're using basically to extrude shape-memory polymer and lock some residual stress

[04:26] And then later, if you try to basically heat it up, the stress will be released, and the filament tends to shrink. on top of a nonshrinkable material,

[04:40] that will bend when you introduce extra heat to it So this FDM simulation is basically trying to simulate what's happening.

[04:56] then we know we can do some sort of a bending actuator. the origami algorithm to compute the local bending angle needed--

[05:11] so that we can generate a G-code with prescribed residual stress And we print everything flat, but once you put it in water,

[05:27] into the Stanford bunny shape. all interesting shapes that can self-assemble from a flat piece, a flat piece when placed on a hot water

[05:41] surface turned into a boat, or a disk turned into a chair. to save packaging and shipping space. But once it's on site, you could assemble them.

[05:55] But in this case, it could assemble itself. And if you're interested in the computational part of it-- so this is a very simple and standard basically origami

[06:07] You simplify a 3D model into a finite number of flat faces. needed at each hinge. from a traditional origami model.

[06:21] In a sense, we need to compensate the width needed need a certain width versus the traditional origami model,

[06:33] So we had to do basically some offset in the algorithm side to take into account the width of each bending region.

[06:46] Depends on what is the bending angle needed. Anyway, so this eventually gave us a software platform where you can basically input 3D model; simplify it into, like I said,

[07:01] flat faces; and calculate the bending angle; And then this is showing once you put it in hot water. So the flat piece can self-fold into this bunny shape.

[07:15] So this is based on this very simple bilayer phenomenon. is related to the residual stress And the residual stress was built in during the printing

[07:31] explore a different physical mechanism which is basically leveraging the local linear shrinkage rate.

[07:43] Basically, we print a single layer. But depends on the print speed or the layer thickness, amount at different local area.

[07:56] Once you heat it up, it morphs into a dome, for example, And you can do computation now to make more arbitrary shape. you might recognize now we can basically analyze

[08:13] So origami shapes are a bit easier to handle geometrically speaking because they're developable. that deal with nondevelopable shapes just

[08:26] like a human face or a very specific mountain terrain It flattens it and generate all the G-code. And you'll see the G-code here.

[08:40] The color indicates the stress or the strain level. And we can also simulate how it would morph into 3D. in this case a flat disk that morphed into a face

[08:55] after you give it extra heat. and then we moved on to look at some other physical phenomenons So this is a representative one that I love a lot.

[09:09] called differential swelling tuned by the surface pattern. You put it in water.

[09:21] It just uniformly grow thicker and wider. But now if we introduce basically sub-millimeter scale grooves on one side of the gel, it

[09:35] will impact the swelling rate and ratio across the thickness. So the side with the groove will end up swell more slowly So if now you basically let it swell in a solvent,

[09:50] in this case a starch gel in water, So essentially, what you get is again showed with the thermoplastic.

[10:02] So this is a differential swelling. So once we get to know a little bit of the physical phenomena, then again, we plug it into some computational pipeline.

[10:17] basically to tell us how the groove are So you can see if our end goal is

[10:30] to get a rose flower, we basically produce a flat stamp in this case with groove pattern on it. to create the surface texture on the gel.

[10:44] And that's what we are showing at the bottom. If we put it in a solvent, in this case, it automatically starts to swell and wraps itself up

[10:56] And we also-- so this phenomenon is really straightforward. You write as a purely structure-modulated morphing

[11:09] powerful because it's applicable, broadly applicable, to soft matter that could swell in a solvent per se.

[11:21] It could be, for example, to some extent, your dumpling cover It also could be some sort of elastomer or silicone

[11:33] sheets being swollen in a polar organic solvent. So we showed in a paper that we can even use So basically, in this case, you can make a grooved flat pasta

[11:49] IKEA idea save a lot of packaging space for food. But once you start to cook them, they take on different shapes, But obviously, this could go beyond food.

[12:05] you could make basically gel-based robots with this. So I'd like to go a little bit deeper into this mechanism.

[12:17] Again, I said the groove is a really unique feature that So initially, you get a flat piece into water. The right is simulation in this video.

[12:31] So you can see it started to bend towards the groove side. And actually, now if you take it and go flat-- now It's a very dynamic phenomenon going on.

[12:48] And interestingly, for a moment, if you ignore the physics, to the pattern of the groove. So it turns out the width of the groove, the gap between grooves,

[13:03] and also the thickness of the groove, will impact how much it bends. So there's a relationship you can derive experimentally.

[13:15] a sense of how basically the curvature will be tunable based So once we get that, we went into our comfortable zone

[13:29] So based on the experiment calculated, we concluded one groove contribute roughly to 12.4 That basically means you get two grooves.

[13:44] You get 24.8 degrees. I contribute to some degree. I added the third one.

[13:59] So that's how you could potentially design flat sheets that can curve into any arbitrary shape. So in graphic world, computational graphic world--

[14:15] you can do this more holistically. you want your flat sheet to morph into,

[14:27] and then automatically derive the locations where you So from that, you could go back to the complicated rose

[14:40] So if you know your target 3D shape like a rose flower, You could then in parallel also derive the scalar field.

[14:53] to the specific location of this flat disk. Then we basically generate a 3D printed mode which function We use that to either mold cast or stamp on a flat sheet

[15:09] Basically, that will eventually morph into the rose flower if you put it into a swollen medium. So with that, you can go more complicated with the morphology.

[15:23] So the one at the bottom is an interesting gripper. You dip it in solvent. Once you take it out, like I said, this dynamic process,

[15:40] That's a moment you can release whatever you are grabbing. Yeah, so basically, you can make a semi-autonomous gripper And this can be applicable for a lot of gels.

[15:57] If you are familiar with microfluidics know PDMS is really basically compatible with the human body.

[16:10] Yeah, so this is the fun application we did. So actually semolina flour-based dough. And once you boil them, they can take on different 3D shapes.

[16:26] On the left, these are showing some finite element simulation. And then you are also seeing some real sample of the pasta.

[16:38] And yeah, we started to-- if you ask why pasta. And it's a very universal phenomenon. Flour-based dough yeast food can swell in your kitchen

[16:53] But yeah, if you also do a simple calculation, about 67% of the space is used to pack air. So if you make pasta that could basically be packed flat,

[17:07] So we actually brought this to Italy. Basically, 1% of the green gas emission So we eventually convinced, yeah, Barilla, the biggest pasta

[17:27] manufacturer maybe in the world, definitely in Europe, So yeah, it's pretty fun. So that's some sort of intelligence

[17:40] I want to share my second point that is that embodied intelligence can be sort of amplified

[17:52] So I want to believe what I just showed are mostly materials. They are thermoplastic. They can do certain things, but they are also quite limited.

[18:07] And some PhDs in my lab through different kinds of projects can combine smart material with some structural intelligence

[18:20] more programmability and controllability in morphing So when I say smart material, yeah, It will shrink.

[18:35] But there's no degree of freedom. There's no sophisticated kind of structural programmability is stiffness changing material.

[18:52] But if you introduce-- in this case, we had an almost invisible heating wire embedded in. it becomes really soft otherwise stiff.

[19:04] recover back to straight if there is no external constraint. So my student started to basically try to combine this

[19:18] smart material into compliant mechanism to do something that's fancier or something that material itself cannot do. make reprogrammable compliant metal structure that

[19:34] has reconfigurability around all six degree of freedoms. as a compliant mechanism. soften or harden certain rod within the compliant mechanism.

[19:53] free that degree of freedom. the degree of freedoms and reconfigurability you want. Again, we know how to address the flexure.

[20:09] the degree of freedom. Rod when it's stiff is a constraint, right? And then we can define which rod is that so-called smart rod that

[20:23] And then we look at the degree of freedom of the entire system So the idea is when it's stiff, you

[20:37] get one degree of freedom status, but when it's soft increase the degree of freedom. This is basically-- this is completely stiff.

[20:52] And this one is up-down rotation unlocked. So you could-- yeah, the palm can basically push inside, And this is left-right rotation unlocked.

[21:06] And then if all of them unlocked, We talk about using this, for example, potentially for rehabilitation or some haptic device.

[21:22] is we introduce a rational design pipeline, basically like an algorithm that allow us to design fully reconfigurable.

[21:37] For example, only X-translation is flexible. This depends on how you, again, arrange nonstiffness-changing rod.

[21:52] The chart is basically showing when you heat it up versus cool it down, you get a drastic force difference So again, the first column shows a device

[22:07] that only has X-translation reconfigurable in terms And the second column is Y-translation programmable, et cetera, et cetera for all the six degrees of freedom.

[22:21] And this is a video showing a bit of the same concept. So again, X-translation now is freed because we Our algorithm told us to heat up.

[22:35] And then this is a Y-translation being freed. It's already a different mode. This is Z-translation being freed.

[22:49] This device really doesn't have any functional meaning. We just wanted to show the generalizability of basically the design method.

[23:01] But I think you get the idea. So once we understand how to design the degree of freedom

[23:13] as we want, we can look at some more specific applications. picture these type of exoskeleton devices can be used, for example, to constrain

[23:27] certain degree of freedom to do like virtual kinetic feedback. For example, if I try to turn a doorknob, while lock the other ones.

[23:41] up the wire, but heat up the wire partially with the PWM so that we can use it to do tunable rehabilitation training.

[23:55] Yeah, and there are perhaps many other applications designers But I really see this project as a way And it's very mechanical engineer-friendly

[24:09] come from an engineering background, And these are some ongoing work. So we're trying to explore some more complex sort

[24:22] of compliant mechanisms that could even have some computation or logical behaviors built in. to my idea of physical cybersecurity.

[24:39] So for this door latch, it's purely mechanical. But you need to introduce three sequential motions to open it So in this case, you have to do, yeah, Z-translation first,

[24:54] and follow with the A-rotation, follow with a Y-translation And this, again, is designed by this inverse pipeline

[25:06] Maybe the left one is a little bit more clear. motion exclusively in order to basically open the door lock.

[25:22] getting into that logical computation you could And if we embed the smart rod, we can reprogram the possible logic you need.

[25:36] you have to do a different sequence. So that's getting into that magical cyber physical security again, this is semi-conceptual.

[25:49] of leveraging this type of reprogrammable compliant mechanism to create multimodal robot. as one groups of lecture that can

[26:05] be softened versus the purple one as the second group. you could reconfigure this robot either be a walker

[26:17] And this is just a simulation. We actually have not been able to implement this This is the walking mode, and this is the swimming mode.

[26:31] At least the compliant mechanism part is correct. We are pretty sure in terms of the swimming or walking we have

[26:43] a lot of optimization to do. Again, I'm just stitching project together to make the point that material in combination with mechanism

[26:56] than pure material-driven system. So this case, we are having-- in this case, the smart material component is the shape-memory alloy

[27:09] in this rubber structure that's rendered in white. And the mechanism part is this white frame and also

[27:22] So it's basically like a bistable mechanism. Basically, we have a 3D-printed frame. And then we basically fix the stretched membrane

[27:39] Because of this elastic energy we introduced into the frame, it can quickly flip to the other side.

[27:52] Then by embedding this shape-memory alloy, leveraging the shape-memory alloy flip the side of this bending actuator per se.

[28:09] And this is-- yeah, this is what it's doing. It quickly flip. you know it's a little bit notoriously slow and pretty

[28:25] But by combining that into a bistable mechanism, into a pretty fast and hopefully more practical one for building robots.

[28:40] has pretty good repeatability in the lab. basically circle back and forth. And these are some of the robots we were able to build.

[28:57] Basically, the body is made of two of those bistable actuator. And at the bottom, you flip the two actuator at the body.

[29:13] And this very tiny body is showing. And once it's in water, it flips its body, Yeah, the wing-- and started to basically pedal.

[29:31] due to the bistable mechanism, you can make a jumping robot. it was pretty magic, because again, the actuator itself is

[29:44] So you could also basically reconfigure the body shape of the robot. you can turn a walker into a roller.

[29:58] If it has to roll down the hill, they form a ball shape. Otherwise, they can just bend up and down to walk itself. Continue with my third point.

[30:14] we discussed this with many colleagues-- how far we can push with basically mechanical system. this morning at Stanford as well.

[30:30] basically robotic arm driven by motors per se, with a very classic robust design. the most precise control, the amount of torque

[30:48] that's more ideal for heavy-duty things. You can easily plug in machine learning or sophisticated control policy to control those robot.

[31:02] are the opposite. They're very weak. Absolutely in the field of soft robots and basically mechanical

[31:16] intelligence, we discuss a lot whether [? they ?] [? are ?] killer app at all for mechanical systems. I think we more or less started to reach the conclusion maybe

[31:28] both mechanical intelligence and computational intelligence Maybe neither could achieve by itself. This is super experimental, exploratory thoughts.

[31:43] I actually can't really think about very convincing But yeah. I mean, those fully programmable robot system.

[31:57] So the question is, can we combine these two? But I'm showing some early experimentation.

[32:09] Maybe give some hints at least from our perspective. So this is an example talking about the design of mesh robot.

[32:21] by a different name, variable-geometry truss, truss robot, or like looped system. and they are connected by node.

[32:36] complex mesh robot like this, in this case, pneumatic linear actuator. You have to have lots of separate control units.

[32:50] So if you want to independently control basically each beam so for a lobster robot like this,

[33:03] So yeah, we didn't perhaps want to do that. So we got quite inspired by-- actually this is during my visit There are some biomechanics people who told me basically,

[33:20] Multiple muscle groups are supposed to be grouped and work together to achieve a task more and think maybe there are ways we could also mechanically group

[33:36] the muscle actuator together. could just have much fewer numbers of module. But allow the mesh robot to do equally sophisticated things.

[33:53] So basically, you can imagine if we need to group them, each node may have to carry multiple air channels. to automatically generate the air channel connection to--

[34:13] and the basic idea is like this. represented by different color. And depends on which group you deflate or inflate,

[34:26] So the design problem indeed is a little bit complex. to be grouped with which one.

[34:38] sequence of all those beams, basically that control policy And yeah, this is one of the example.

[34:50] Again, that lobster designed by the idea I just mentioned. And we also tried to implement a physical version. two tubes in this case.

[35:05] So basically, this is grouped into, I believe, two groups. If you try to do it manually, you

[35:17] have to think about the grouping. Also, each actuator control how much contraction ratio Also, the control sequence, basically the control policy,

[35:32] which group will be triggered at the same time, things like that. Now, if you want your robot to do multiple tasks, What if I wanted to turn around while lowering its body?

[35:46] So it becomes a multi-objective design task. Eventually, we decided a simulator and a manual design. beams in a software system is not really practical.

[36:01] So we started to get into basically optimization method. genetic algorithm to figure out how to do Basically, given all the input design factor,

[36:16] and how to control the inflation, deflation. reinforcement learning to handle similar tasks.

[36:29] And recently, we started to also leverage generative algorithm to come up with the robot design as well. And yeah, this is what we had out of the genetic algorithm.

[36:43] being able to do multiple tasks through the optimization method So it depends on basically the channel grouping and also

[36:57] You could let the robot do different things while not changing the grouping obviously. tilting, lowering the body, and things like that.

[37:15] such complicated mesh anyway? Maybe like a single unit. department at Stanford, they have some really cool robot

[37:30] made of one or two unit, basically maybe like a 10 truss. partially because we can't do it with the algorithm, but also partially because we were interested in understanding

[37:44] some unique design spaces. For example, if you think about your entire helmet is a robot, can you possibly morph the shape of the helmet

[37:58] Maybe depends on the scenario when you wear it, [? wear ?] the same helmet can morph So this itself is a morphing-- it's again a multi-objective--

[38:18] basically task you can frame it. at the bottom, we're just showing So this is the simulation result. We actually build one.

[38:32] We were tracking basically how precise each node could match our simulation or not. you could really start to think about all complex robots

[38:48] Even daily products you can build with it. So we are speculating some more complex things,

[39:00] but algorithmically buildable hardware-wise, not sure. Basically, like a morphing mesh bed or automatically adjusts the patient's body

[39:16] posture because they get tired being fixed in one pose. you could really physically render different physical objects in a mixed reality environment.

[39:32] Or in this case, yeah, basically give people Say you are playing a game, you turn into a bird, how that physical mesh can morph.

[39:45] Your center of gravity will be shifted are carrying as well. OK, so my last point I want to articulate, yeah,

[40:00] why embodied intelligence is interesting, what's it good for. I guess I didn't give an answer, but these are definitely My lab was also very interested in looking at basically

[40:18] how purely ambient energy-powered system Imagine a field robot deployed massively into the field.

[40:34] want electronics to be part of it, why maybe you want minimum intelligence, but be able to basically execute the task as it's intended to.

[40:51] ecological physical AI or ecological morphing matter. some mechanically intelligent robot,

[41:03] So the environment give us the energy, whether it's sun, or again moisture fluctuation, And you design a physical system that could smartly

[41:20] some sort of a mechanical computational unit that make decisions in terms of what you use those energy to do So that's a framework we started to develop.

[41:37] And this is not published yet, but it represents So how you perceive basically natural stimuli And how do you build mechanical devices that rationally filter

[41:53] And how you couple that with morphing materials and structure So for example, like a facade piece, could respond to very specific moisture and light condition,

[42:12] automatic open and close to tune the light permission on your window, or how a system built to basically harvest energy and do robotic work in the field to manage a garden.

[42:25] I didn't mean to basically let the audience fully digest it, but just visually give you an impression where we are going. So you could use the lens of embodied intelligence

[42:39] to interpret how some of the natural systems work. Not getting into details.

[42:51] So here is an example in terms of again purely passive little but play a role in natural condition. So this project is inspired by a seed from nature.

[43:06] Basically, it has a coiled body that's responsive to moisture. The coil will unwind. And that unwinding generates a rotational motion

[43:20] downwards to push the tip into the soil. Again, the coil unwind. The whole thing rotate and then push the tip into the ground.

[43:35] So this is also a little bit like your power screw. to try to make a biomimetic version of that.

[43:48] So on the right, this is made in a lab. a wooden veneer, a white oak or maple wood veneer. you can carry the target seed you want.

[44:02] So actually, right now, we're working with the CAL FIRE Department and also over the East Coast with the Cornell Forest, try to actually engineer seed carriers for specific tree

[44:17] So the idea is that you could carry many, many of those. They are made of cheap wood veneers. And drop them onto the ground, carry the target payload.

[44:31] Drilling into the ground is very helpful to improve the germination of the seeds you are interested in. And this is a video showing basically

[44:46] like an earlier field test we did at Pittsburgh in this one. And again, it's a really early stage test, But we were able to show we can drop them onto the ground

[45:02] and response to the rain in the afternoon and self-drill So through the project, we also got and we learned basically reforestation again is still

[45:16] people basically have to do seedlings. They manually go up to the mountains, put the seedlings in. if you have to go to a sharp slopes to do seeding.

[45:33] So it makes a lot more sense for us germination of tree seeds. looking into how we could possibly do things better

[45:49] I'm not sure if you guys noticed the one in natural system So later, we started to add more tails, play with the improvement of the [? torque, ?]

[46:03] There are a lot of basically engineering optimization work going on in the lab right now. Basically, we're showing we got some initial success

[46:19] And we also thought about how we could engineer those not so smart carrier to be a little bit smarter in this case to it and the wireless communication as well.

[46:37] them to do things at a little bit of a networked scale. This is the earlier demo to show you could basically

[46:49] equip a antenna to do wireless communication. The vision is you will be able to monitor some conditions, maybe underground condition,

[47:01] and then send those information back to a drone that flies by later on. So I have this slide.

[47:15] helpful for reforestation. for fertilizer, basically precision agriculture to save fertilizer uses.

[47:30] into space, so a lot of the earlier study of seed because they were also interested in basically environmental-friendly, completely like energy-free,

[47:45] can be used to do soil sampling in outer space, is harvesting the moisture fluctuation, can be used, for example, to harvest thermal fluctuation,

[48:02] the moisture-sensitive actuator into So yeah, there could be some really cool applications

[48:14] And that maybe is also a killer app of mechanical computer. very short-- again follow the same threads of thoughts,

[48:29] how mechanical computer can be a good use in outdoor. envisioning a fully automated garden,

[48:41] from seeding to watering to fertilizing for extreme weather and purely with mechanical systems.

[48:55] And we see them as robots because they have-- They also have actuation. multiple stimuli-responsive pump and valve

[49:11] The pump is a way to basically harvest ambient energy and turn them into pneumatic energy, basically compressed So for example, this video shows how this component

[49:28] is designed to harvest the thermal fluctuation in the air and accumulate compressed air into a chamber. So basically, it's a thermally responsive pump, right?

[49:41] What it means is that when the moisture comes, That's why it's a moisture-sensitive valve.

[49:53] It's done by a very classic mechanical mechanism called The idea is that when moisture comes, get swollen and then basically close the kink valve

[50:08] by literally kinking the tube. And eventually, we are able to construct harvesting devices and valves.

[50:23] So you could harvest thermal fluctuation, You can also harvest kinetic energy such as wind and hydro all pipeline them into a compressed air.

[50:36] respond to different environmental stimuli as well. to construct pneumatic circuits that take account In this case, this is getting very complex.

[50:52] If you are not, I guess, super into pneumatic actuators-- But example make life easier. like a dispenser that automatically get

[51:07] in the environment is right. And then after a certain amount of time, And then after a certain amount of time--

[51:22] And then when you detect the right wind condition will come out to protect the plants. There are a lot of if and when and and condition.

[51:38] So basically, there are a lot of logic being programmed about this kind of fully self-regulated plant device

[51:50] So to wrap up my talk, we discussed over lunch, hey, maybe the future super powerful devices

[52:04] both from the machine and also from the material intelligence. Or you can think about it as a system that's hybrid.

[52:17] And also, you could have biological mechanisms or biological materials, living or dead, being part of it At least that's the space, the intellectual space,

[52:32] my lab is tinkering around at the moment. [APPLAUSE]

[52:44] Any questions from the audience? So for the materials, you said [INAUDIBLE] properties

[52:57] of diffusion to whatever's soluble [INAUDIBLE] solution Yeah, that, for example.

[53:09] So when we do finite element analysis on those structure, we basically plug in all those factors. We do numerical simulation.

[53:24] I would measure the diffusion ratio, the change in modulus, and just let the numerical simulation does its magic per So it will automatically take account-- basically,

[53:37] with or without the groove. play a really important role. But when we do that very simplified groove geometry

[53:51] matching to the contributed bending angle per groove, that, so I only have half of the story.

[54:03] know play a very dominant role but not the absolute role. as well. but the computation is really lightweight

[54:17] We do experimental-based, data-driven process to understand how basically the groove contribute to the angle. We do the initial guess.

[54:32] to optimize the groove pattern. that's based on more precise physics-based simulations. [INAUDIBLE] geometric properties are dominant enough

[54:46] Yeah, yeah. you have to go back to more sophisticated multiphysics model

[54:58] Yeah, because it is a complex problem, right? As you can imagine, most of the cases, it gets saggy.

[55:10] So the modulus changes, as you can imagine, will play a role. yeah, yeah, the thickness will change the bending stiffness.

[55:22] Yeah. often challenging when we deal with morphing materials because none of them are like black and white.

[55:37] everything is so calculatable. But here everything is so nonlinear. [INAUDIBLE] really, really inspiring to see all the work

[55:55] that you have done, especially with ecology and robotics I'm really curious to hear how you integrated into future products that people use

[56:12] lots of really hard hardware. And so I'm just curious what the future is in your mind for that. As you can see, my approach of thinking

[56:27] are quite, I want to say, to some extent, So ours always try to go some [CHUCKLES] semi-niche area.

[56:41] Yeah, like a seeding seeder or pasta being a [INAUDIBLE] food At least it's one perspective I see how nonconventional robots get into products, even

[56:54] But there are some mainstream applications for soft robots Steve showed me earlier.

[57:06] for more rigid robotic arm. And you can handle much delicate soft objects like food or jellyfish or whatever.

[57:20] And my student also started to look a little bit into how you can do a little bit more, I guess, So he was working on-- he was showing me

[57:32] But you could-- a soft skin mounted on a rigid hand, If you don't do anything, it's just rigid.

[57:45] But then you can inflate bubbles and then becomes very compliant. Now you can grab a delicate cake, cupcakes or whatever. So now you can grab very smooth object.

[57:57] I see this is a little bit relevant to what how you leverage the smart material aspect that but use those to augment some of the more precise,

[58:14] I feel this could be one possible direction that's Yeah. Can you say more about the trusses examples

[58:30] Where does that work going, or where is it going? So far, it hasn't gone anywhere because my student basically

[58:42] worked on the sort of the whole concept, the mesh And we tried to propose to a hospital, local hospital, The proposal didn't get through, so we haven't got to work on it.

[58:58] Actually, we are right now at much better position to work on a system like this than two years back. And we also discussed-- so the student thesis--

[59:10] so the algorithm is not only applicable for this type So the algorithm can be generalizable, for example, other closed loop graph-based robot designs,

[59:23] We actually think tensegrity robot is likely more practical than the mesh, because this mesh doesn't have a recovery force to let it recover back to its undeformed state, basically.

[59:37] algorithms. Yeah. collection.

[59:50] what was the particular thing that your solution addressed? Was it a particular difficulty of getting quality

[1:00:04] You know what I'm getting at? This is a conceptual image. we're right now having an effort to write a vision paper.

[1:00:21] So some of the diagrams related to ecological things I just showed is part of the paper, like those things I mentioned. I don't know how to-- yeah, it's part of this bigger framework.

[1:00:35] robotic system. And we've been brainstorming what are the use cases. Environmental DNA just came into the conversation

[1:00:47] because we were thinking about the residues left over So we may not need a robot that does very precise navigation

[1:01:00] that hops or roll around semi-random, So we're thinking this could be just a way for them.

[1:01:13] and collect DNA as long as we know roughly [? It's ?] where we are thinking. But no exact project at this moment is going on.

[1:01:33] [CHUCKLES] Yeah, and also this image is showing the flying machine.

[1:01:48] is working on, basically like dandelion and milkweed-inspired use it to collect environmental DNAs in the midair

[1:02:02] And the rollers will be on the ground, basically. has done so many different source for actuation,

[1:02:17] [INAUDIBLE] temperature. or what are some other potential modality of actuation that you will be excited to explore in the future?

[1:02:35] At this point, I feel like motor is awesome. there's no single ideal one. Most of them are weak.

[1:02:50] And it really has to be application-specific. Believe it or not, the wood actuator, the seed carrier, indeed is one of the very stiff one because even when it's wet,

[1:03:02] the coil, the material itself has still about 5 to 10 [? megabyte ?] or 10 [? megabyte, ?] like this soft I think it really depends on the application.

[1:03:17] but it's going on in the lab, through a defense-funded project, we are working on biohybrid.

[1:03:29] So those would be like a real human skeletal muscle. being like self-heal and even self-pieceable, but obviously And we are also working on--

[1:03:44] fluidic-driven actuator. So those-- yeah, for example, electrostatic force

[1:03:56] So those type of actuators are at least electronically You drive a fluid. to our vision of hybrid system.

[1:04:11] of, in this case, hydraulic actuation. Any last questions before we head to the [INAUDIBLE]?

[1:04:26] I was going to ask one more, another thing in terms in the environment for various species in terms of adapting quickly enough because the challenge

[1:04:40] But we're going there very quickly. like, if there's a weak point in a plant or something where it needs like some sort of help in delivering seeds,

[1:04:55] or if there's some weak spot in the chain of the [INAUDIBLE] of that species, something like this could provide. Yeah, yeah, yeah.

[1:05:08] [INAUDIBLE] could provide some solution that could allow for a weak spot in just the time the species can evolve [? in ?] I mean--

[1:05:22] I'm not sure it is right now. I mean, there are, for example, seagrasses in California, So they change color because the weather change.

[1:05:37] the condition. I can imagine some sort of color changing variable scheme for the plants maybe to tune that.

[1:05:50] Yeah, yeah, yeah. All those are very interesting context.

[1:06:02] to get into any of those. We are getting a little bit more knowledgeable about like, in ocean.

[1:06:16] Haven't got into understanding the problem space too much yet. Let's thank Professor Yao for the great talk. [APPLAUSE]