An Interview with Dhanush Radhakrishnan

Marwa ElDiwiny

Dhanush Radhakrishnan

Marwa ElDiwiny: Please tell us the story of Clone. How did it get started?

Dhanush Radhakrishnan: The prehistory of Clone started almost 9 years ago now, in 2014, when my cofounder Łukasz [Koźlik] had seen a picture of an artificial muscles-actuated leg in a textbook and was very inspired to build his own android. He was relentlessly iterating over various classes of artificial muscles, as well as making prototypes of hands, which he saw as the hardest problem to solve to make entire biomimetic androids. Not only is it the hardest problem, but also it is the most relevant problem to making these things useful.

When I had taken notice of him, and my interest at the time, I was deep in machine learning, and really wanted to put a brain on this thing, productize it, take it to market, scale up manufacturing, etc. We hit it off superfast. We thought the exact same way about how we wanted these things to work.

We did not want to compromise anywhere, to set the vision as making these biomimetic superintelligent androids, and not settle for anything less than that. To go as one-to-one with the human body as possible. Łukasz was leveraging principles of biomechanics that he was studying for years to make these prototypes. For my own research, looking at various classes of artificial muscles, it was clear that there was nothing better than McKibben's.

While I was looking for something that could be an electric, purely electrically actuated fiber or something like that, there was nothing that really came close to the performance of an actuator with a rubber tube essentially surrounded by some kind of fiber-braided sleeve. This is just the basis of the actuator, and from there you can take it very far, much farther than what people would imagine.

I think the story of Clone is about the relentlessness of making possible what other people do not think is possible in every dimension. Everything we make is custom: custom bones, custom valves, custom muscles, custom sensing, like a custom pump motor. It was done out of need, not because we wanted to do everything custom. We cannot get muscles on the market. We bought Festo muscles, but they would not fit inside a hand. They are pneumatic and would require a bulk compressor. It is just not going to work. You need to make your own muscles. When you make the muscle, you realize that there are issues with the fluidic actuator, so you need to make it better in many dimensions.

You need to improve the braided sleeve, you need to improve the way that you manufacture it, so it can be done very quickly and cheaply, and can be scaled up. The story of Clone is “How do we get biomimetic androids?” Well, you need to figure out a million problems. And when we started the company, I was lucky enough to have a cofounder like Łukasz who already has shown that relentlessness.

Marwa ElDiwiny: You have said that you do not want to hire a traditional control system professional, and to use neural network work for control? Can you explain this statement?

Dhanush Radhakrishnan: We are not dogmatic about anything. People who come from the world of motor-actuated robots tend to not understand the countless ways in which you need to innovate your way out of problems to make androids the way we do. And there is too much of a comfort almost in getting motors off their shelf, getting batteries off their shelf, making robots the way everyone else does, that on the hardware side, at least, you do not innovate your way out of these problems.

But also on the control side, it turns out that when you are making such a high-dimensional system to begin with, it is so difficult to control classically, the number of nonlinearities at the joint level, the muscle level, the tendon level. They accumulate so quickly, and it is silly to even attempt to classically control it and pretend like you could deliver a product to customers where the robot is doing real work, and you are classically controlling the whole time.

That is not going to be tenable, so why even pretend like it is something you should do in the early days? It helps you on the hardware side too, because if you can make the assumption that it is not tenable, then you can start making much easier trade-offs on the hardware side where you can just say, “Hey, I don't care about these non-linearities. It doesn't matter at the end of the day, because I'm going to be controlling you with a neural network, so I can optimize my hardware design in all these other dimensions that make the hardware better in terms of power density, strength, speed, etc.”

Marwa ElDiwiny: What is advantageous about the design at Clone? And how do you decrease the noise?

Dhanush Radhakrishnan: One thing about the noise is we are using hydraulic muscles, not air muscles. This is incompressible flow. We do not need an air compressor, a noisy compressor in the first place. Łukasz was working with whatever he had around him. A lot of the parts were not great parts. He was just being resourceful and using washing valves that he pulled valves out of a washing machine, things like this.

Obviously, if you use components like that, there is going to be noise. It is very steampunk and not exactly optimized for a real product. As we went about optimizing the design for a real product, you can get rid of the noise in the valve, you can get rid of the noise in the pump and motor. This is even without any acoustic insulation. You can imagine that with acoustic insulation, noise goes to virtually zero. I mean in the most recent videos of the hand, the noise that you can still hear in the videos is squeakiness in the servo valves, and we are not using those anymore Our new valves are completely silent.

Marwa ElDiwiny: What is still challenging for you in Clone? What is the thing in which you are still trying to push the bar of the design?

Dhanush Radhakrishnan: Oftentimes, when you solve one problem, it creates other problems. What is left for us is compactifying our energy supply. We have been thinking about energy from day 1. It is not like we do not have ideas, we do not have solutions, we have plenty of solutions. The best solution we have is one where this is going to be a battery-powered robot, an electrohydraulic robot, which will have enough energy for the entire android to run for your entire waking hours without needing a recharge. We are not interested in making an android that is going to run like Boston Dynamics' Atlas for an hour and then needing a recharge.

Marwa ElDiwiny: Can you elaborate a bit about the power consumption?

Dhanush Radhakrishnan: There are a few major differences between the Atlas and our robot. Atlas is the other well-known hydraulically actuated robot, but they use heavier piston cylinders. It is a much bulkier robot. The weight is not distributed evenly. There are a lot of ways in which they need to expend a lot more energy to control their robot. And then we do have an ace up our sleeve that I cannot talk about here, but it is going to multiply our energy density by several times. We are not too concerned about the power supply. It was annoying to get the power consumption down in the valve, but it turned out that by switching the valve to our most recent valve, we brought down both power consumption and noise together. This has been the biggest breakthrough in the company yet.

Marwa ElDiwiny: I am curious about the next steps. Do you plan to design a leg with the same design as the arm? What are next steps for Clone that could change the situation for people who would need a prosthetic arm or leg?

Dhanush Radhakrishnan: We are pursuing an android first before we end up making prosthetics. We are very firmly in the robotics category right now as a company. In terms of what is next, we are first focused on selling some hands to research scientists who have been wanting them for some time now. The biggest blocker to that, most recently, has been needing to finalize the valve, so that when they purchase a hand, they do not have to redo a lot of the work they did when they purchased their second hand, which would use a different valve. Many of these user scientists—world-class laboratories from industry and academia—who want our hands, have been waiting on a finalized design.

Up until 2023, since we started the company, the focus was on durability. Getting the hand durable enough so that you can run it without any human interventions. You can run a learning algorithm in the real world, and let the hand continue learning to solve some manipulation task without any human intervention. That is the thing we got to by 2023. We are very happy about that, it makes it ready to be sold.

But then the final thing was just that researchers did not want to have to deal with the hand that was going to change in design very quickly, iteration after iteration. We have been updating the hardware quickly. We are just about there, where we are finalizing the valve design. It is miniaturized; it is consuming very little power. It fits in the vein, opening up the entire torso volume for the energy supply.

Marwa ElDiwiny: I am going to ask you about the actuation here, because I think what Clone is doing here is very interesting. Do you have any plans to combine different actuation techniques? How do you see the design that you and the company developed could be deployed for legged robots or other forms of what we have in robots now? Do you think you should or should not use hybrid-actuation techniques?

Dhanush Radhakrishnan: There is no need for a hybrid actuation technique. Our design is simple. Once you have the experience of making it a thousand times, which we have, it is straightforward, and it does scale well to the entire musculoskeletal android. To answer your previous question in full, the immediate next step is to sell hands to research scientist customers. Very soon after that, we will be scaling up the musculoskeletal design to the torso and then the full biped.

In terms of actuation, you have a skeleton and then you put muscles on it. This is different from every other way. Nobody else makes robots like this. Everyone makes a rigid exoskeleton and then you stuff as many batteries and motors inside as you can. That is not how you have to make robots. You can just start with the human skeleton. You can just put your actuators over it, and then wrap it up in skin, and wow, you have an android. And that is the approach we are going for, and it scales nicely.

Marwa ElDiwiny: I am curious about iteration of design. I am curious whether you can share something that was counter-intuitive to you and Łukasz that the design should work, but it did not work in the early stages? Maybe it was a learning lesson for you to come to what you have today?

Dhanush Radhakrishnan: Not really, because, in the first few years Łukasz was working on this, he had some more different-looking hand designs. Once he settled in on McKibben's and the musculoskeletal just looking design, it was straightforward from there. We do not really design, we just copy human anatomy, we dissect humans in the dissection room. We look at what is inside, and we look at the textbook, and we just copy it.

Marwa ElDiwiny: Do you think that in robotics, there is something missing, not enough work invested in something similar to what you do?

Dhanush Radhakrishnan: From our perspective, absolutely. The reason you could not design this way is because nonlinear control was just not feasible, and you could not do high-dimensional control a few years back. It is only possible since 2021 onward, I would say, 2022 onward, where high-dimensional control started to become feasible. If you look back at the Shadow Robot Company from some years ago, maybe almost a decade ago now, they had a muscle-actuated hand briefly, but you could not control it. If you wanted to teleoperate a robot back then, because you could not do autonomous operation, neural networks were nowhere near good enough.

If you want to teleoperate, it is much easier to control some linear motors than it is to control these nonlinear muscles. And that was l the problem. Łukasz had this certainty, even without knowing where it was going to go, just knowing, “Hey, AI is going to get better and better and so it should be fine.” And I jumped in as I knew, “Hey, AI is good enough now. We can train very large world models that can learn all kinds of rich information about reality that'll let us control these very high dimensional systems.”

Marwa ElDiwiny: What is your vision for remote teleoperation versus reinforcement learning-based autonomy?

Dhanush Radhakrishnan: It is not just reinforcement learning (RL), or teleoperation, but RL is a critical part, I think, of autonomous operation here. If you had asked me what our biggest disadvantage as a company was, it would definitely be that in the early days, without neural networks, it was hard to control this thing and hard to put out demonstrations. That is why, despite having great hardware, it took us so long to put out a single manipulation demonstration. Even the types of learning problems I had solved in the past were so low dimensional compared with this problem. The learning problems I had solved in the past were not actually being applied on hardware.

We finally put together a teleoperation demonstration with proportional-integral-derivative (PID)-type control, that is still rather finicky to work with. But teleoperation is, I think, going to be more short term than people realize. It is just not fun to scale if you do not have a nice rig where it is fun, the fine control is good, you have haptic feedback, so it is a faster learning curve to use the teleoperation rig. For us to do haptic feedback, we would have to make some kind of glove, and we would just be spreading ourselves way too thin at that point.

For us it is going to be limited teleoperation, trying to get some number of teleoperation play experience in virtual reality (VR) and, with that, augment the data as best as we can. Use some diffusion models to take a single experience and then change up the environment on that experience a thousand times over.

Then, also, you have the teleoperated experience. You can turn offline RL on that operated demonstration and let the robot trial and error that same type of short-horizon task over and over until it is just getting more active offline data, besides your motor data of course, to solve that task. At the end of the day, if you can see the trend on the language side of things, the aim is to get these robots working autonomously in the real world. We are talking about something like billions of neurons that humans have.

Getting the analog of that, it is a ton of data that is required, a ton of experience. Where the story of language, of course, has been deep learning, Stochastic Gradient Descent on large amounts of data, that is what works, and that is what lets you see the magical results you see in on the language side of things from deep learning. But of course you are working with the copious amounts of Internet data.

I think in 2022 what changed was that people started to figure out, “Hey, you can leverage video data of human hands on the Internet,” the millions of hours of video data to start leveraging it toward motor control. The way they were doing that was something like making large pretrained visual representations, and then separating that from low-level control policies. I think that going forward, that will all be end-to-end, where you were just pretraining entire behavior representation, rather than visual representations. To do that, of course, you all need a ton of video data, but you all also need some sensory motor data.

The video data you can get from the Internet, you can get from the robot just standing there, and watching you do stuff. Past that, though, you are still going to need some threshold amount of sensory motor data. The first part of that you need to get from teleoperation, so it is high quality, and after that you can get it through trial and error. You can get it from data augmentations, which you can do in various ways these days.

Marwa ElDiwiny: You mentioned that you have the same exact design as human, and I want to ask about the material aspect. Do you ever consider the designing material that could be as a sensor and actuator and resemble the human muscles?

Dhanush Radhakrishnan: I would not talk about it as just material. There are certain parts of the body where that makes sense. You can think about ligaments being sort of elastomeric, and if you can maybe make your robotic ligaments elastomeric as well, that makes a lot of sense. But in terms of the muscle, it is more about the contraction properties of the muscle that you are trying to replicate. With human skeletal muscle, what you are trying to replicate is this amazing contraction ratio for a very low energy input, and, of course, the shape of it.

The material itself is more about the property. If the material has a property where you can run a small current through some magic material fiber or something, and then the fiber contracts 60%, it sounds like a magic material that would be great to replace our muscles with.

The muscles we are using give you the best properties, the best contraction ratio, and the strength because of the contraction ratio. Speed can actuate quickly, with very fast response times, extremely low cost to make, and follows the shape of skeletal muscle approximately very linear. It is more about approximating all these various properties as close as possible, rather than saying it feels like what a human muscle feels like.

In terms of what Will Jackson has said about not getting anywhere close to a human skeletal muscle, I thought I would address that. In the forearm, at least, I think human muscles are contracting around 20% to 30% and ours do the same, 30% with no load, over 30% with a 1-kg load.

These are great muscles already, but in some parts of the body, I think in the leg, especially, your human skeletal muscle contracts up to 60%, I think. There is some gap there. For all intents and purposes, you will not notice it in our design. You will not be able to tell, “Oh, it's contracting 20% less. It's not a skeletal muscle.” We have one more improvement we can make on our muscle that can achieve the human skeletal muscle; it is the same type of muscle as what we would make, but with a slight modification, and it would be able to reach I think those 50% to 60% contraction ratios that some parts of the body, some muscle regions in the body, contract at.

Marwa ElDiwiny: Could you embed everything in the arm itself, so you would not need an external device?

Dhanush Radhakrishnan: For a prosthetic? That would be nice, but we do not know how to do that just yet. We can make it compact. What we have been focused on making are all the powering systems and everything compact enough to fit in the torso for an entire android to operate for a whole working day, for example, without needing recharges, and to do that wirelessly. That itself is hard. We are getting close to that.

I think upon making that work, chances are some of the know-how I mentioned that we have, that I cannot talk about in too much detail, including how we make this energy dense enough to operate for several times longer than Atlas, for example, will be critical in making a prosthetic in the future that will have a very light and portable powering system that is just so small that you do not notice it as a wearer.

Marwa ElDiwiny: Do you think there are any beliefs in robotics or the robotics community that should be changed or challenged from your perspective?

Dhanush Radhakrishnan: I am very antibeliefs. I think beliefs do not mean too much. You either know, or you do not know, and if you really want to make this biomimetic android, you are just going to invent your way out of problems, and out of dead ends. If you get dogmatic in any way, you are already killing yourself and preventing yourself from getting to where you want. If you want a biomimetic android, you need to be willing to trade pieces as often as needed until you get to the right solution.

We had to do that so many times over before we got to where we are now. There is probably a couple more trades for us to make before we have the full android next year. But we see the path in front of us, and we have cleared most of the way. I think, you will likely see more of our design out there as soon as we put out our first torso, our first android, and people can see the delta in performance of having an actual two-handed full android, able to do both strong tasks and dexterous tasks, and walk around without needing to recharge often. I think this will blow minds, and it is likely that we will end up seeing more designs like ours.

Marwa ElDiwiny: Can you share some information about the timing, the time operation, and charging?

Dhanush Radhakrishnan: It is going to depend on what the robot is doing. If you are doing back flips continuously, they are very different from just cutting some vegetables, which requires very little energy. If we are talking about say, an android on an assembly line, working double shifts, it should be able to do this, with the power bonuses we plan to add to the android that are not in it right now. We are shooting for 16 hours at least of continuous operation without recharge.

Marwa ElDiwiny: What are the things that you still aspire to have in Clone? What kind of crazy ideas do you still want to have?

Dhanush Radhakrishnan: That is always a fun question. I think we are very much the android company. We do not have to make any other products other than the android for a very long time. With the musculoskeletal type of technology, if you have any rigid skeleton, any skeleton based on anything, any biological creature, you should be able to put muscles on it, and actuate it, and make it work, whether it's an animal or something else. I imagine that giant androids for construction would be interesting to see. Maybe the crane is outdated as a human-operated automated tool, or a human-operated tool for automation. Maybe a giant android to replace a crane would be interesting to see.