IHMC’s Atlas Learning Crane Kick, Will Destroy Competition at DRC Finals
IHMC Teaches Atlas to Walk Like a Human
IHMC’s ATLAS Robot Learning to Do Some Chores
The robotics group at the Institute for Human & Machine Cognition (IHMC) in Pensacola, Fla., has an enormous amount of experience with walking robots. They came in second at the DARPA Robotics Challenge with their Running Man Atlas, one of just three teams to score a perfect 8 out of 8, and they’ve continued to advance bipedal locomotion using both Atlas and NASA’s Valkyrie. We write about their research all the time—just a few months ago, they taught Atlas to walk with straight legs, much like a human does.
Humans set a very high standard for bipedal mobility. We’re well designed for it in both hardware and software, and we can do some absolutely amazing things.
Getting robots to do the same kinds of things that humans can is an intimidating challenge, requiring both complex hardware and innovative software working together. It’s hard to say which is more difficult, but you can get a sense of the current state of things when you compare what humanoid robots are capable of in simulation with what happens when you try to run that same software on the real robot. There are many reasons why this is the case, but the fundamental problem (or, one of them) is that we’re trying to get robots to do what humans can do without access to hardware that’s anywhere near as good. Biological muscles are incredible things, and while electric or hydraulic actuators can compete with them on things like speed and torque, there’s no way of getting around the fact that the hardware is far too bulky to be able to replicate the densely packed muscle groups that let humans do what they do.
All full-size bipedal robots have this problem, including Atlas and Valkyrie, but IHMC has been using them anyway, since there really aren’t all that many alternative platforms that are accessible for researchers. It’s gotten to a point where IHMC wants to do more than these platforms are capable of, so they’ve decided to do the sensible thing and build their own humanoid from scratch. And if you’re going to do that, why not crank everything up to 11 and decide to build not just any humanoid robot, but a humanoid robot gymnast?
Images: IHMC Nadia 3D-printed plastic mockup in different poses (clockwise from left): foot up, stretching, kneeling, curled up in a ball, and sitting.
This is a 3D-printed mockup of IHMC’s new humanoid. It’s named Nadia, after famed gymnast Nadia Comăneci, and is being funded by the Office of Naval Research (ONR). The mockup contains (almost) all of the components that will be on the real robot (plus a motorcycle helmet that may or may not be the robot’s real head), and working with the design this way helps IHMC to make sure that everything fits together properly and that the actuators and joints provide the flexibility and range of motion that they’re looking for. Nadia will be powered by hydraulics, using Moog’s Integrated Smart Actuators (ISAs), originally developed in collaboration with IIT for the HyQ quadruped.
The ISAs are 3D-printed out of titanium, and are completely self-contained. They have a power-to-weight ratio substantially better than human muscles (if you don’t count the rest of the hydraulic system), although they’re still bulky enough that designing joints around them can be tricky. It’s especially tricky since Nadia (a gymnast, remember) needs to have the same range of motion that a human does, with the ability to do things like squat, crawl, curl up into a ball, take a yoga class, or do whatever else humans are capable of.
A robot with this level of flexibility and the kinds of dynamic motions promised by those ISAs sounds very ambitious, but if anyone can do it, it’s IHMC. And for more about Nadia, we spoke with Robert Griffin, a research scientist at IHMC, via email.
IEEE Spectrum: Why is now the right time to develop a new humanoid robot from scratch?
Robert Griffin: First, we at IHMC have a need for some kind of new humanoid. Our algorithms are really pushing the physical limits of our Atlas and Valkyrie, so beyond just their capability limitations, if we want to do any more research on locomotion and control, we need some kind of new hardware. Second is lack of availability of alternatives. There is Cassie as an option, which looks like a good robot, but it’s not a humanoid, strictly speaking. While Digit has an upper body, I haven’t seen anything about sale price, or research availability for that platform.
Third, we think we have the capability of making a really good robot, and would like to have a next generation humanoid in-house. We’ve done a lot of work with other people’s platforms, and while they are typically very well done, we see a lot of advantages in doing research with our own robot. Fourth, that we have access to a lot of the enabling technologies now. The ISA looks to be a phenomenal piece of actuation hardware. Perception sensors are becoming commodity items now because of the autonomous car industry, and IMUs can be had for cheap because of cellphones. With all of these key pieces in play, we think we can do something incredible with our design.
What applications would Nadia be ideal for, and what are ONR’s goals with funding this research?
The primary application is to have a robot that can operate in indoor environments. ONR itself has been interested in having robots work alongside people for years, with projects like the SAFFiR program for shipboard firefighting to robots performing maintenance tasks to robots doing room-to-room search based tasks, similar to inspection. However, an indoor environment represents a really hard locomotion challenge, with stairs, ladders, cluttered floors, etc. that we think legs are perfect for. This can be solved with UAVs, which are a lot easier and cheaper than a legged robot, but then you throw a door into the mix, or require the robot to do something physical, and your back to needing some kind of ground-based platform just for load capacity.
We’ve seen some fairly quick, fairly dynamic robots that use electric actuators. Why go with hydraulic actuators for Nadia instead?
Electric actuators are a fantastic solution for quadrupeds, particularly smaller ones. One of the common answers has been to use the large diameter, lightly geared electric actuators to get high speeds and fairly high torques. However, because of the diameter of these actuators, packaging them for a 6-DoF leg is really challenging. Agility did a great job with the packaging of Cassie, but I think it’s worth noting that it’s a 5-DoF leg, rather than 6. While 5 is more than sufficient for many tasks, there are some research problems that we’re interested in where we think you need that extra DoF, such as the work that we did in balancing on line contacts.
Additionally, we’re looking at a robot that’s probably a little heavier than those by Agility Robotics, so you need a higher power density. We just haven’t seen an electric motor that can beat a well-packaged hydraulic one in the ways we need. We think that’s one of the reasons you still see these really, really dynamic, large-scale robots like the new Atlas using hydraulics. With the right hydraulic power units, valves, etc., you can source a ton of power really quickly for some really explosive movements. We’re hoping to be able to demonstrate some of this speed and power with Nadia, while also maintaining a really high range of motion.
To what extent do you hope that Nadia will improve upon the capabilities of robots like Atlas and Valkyrie when it comes to relatively “simple” tasks that these larger and less dynamic robots still have trouble with, like reliable walking over rough terrain, disturbance rejection while moving, that sort of thing?
The faster you can step, the (literally) exponentially easier the balance problem becomes. While we’ve had some really good results lately with Atlas walking over cinder block fields, it will become significantly more robust when we decrease our swing times. That means that we can respond significantly better to uncertainty and disturbances and such. Our goal is to get to where we can walk over these types of obstacles at a slow-human speed.
I think the biggest improvements will show up in manipulation tasks, though. Atlas and Valkyrie are really slow at them, partly because their workspaces are so small. We hope that by having a big workspace, we’ll be able to do a lot more practical tasks.
How much of humanoid robot development is a software problem and how much is a hardware problem, and where will Nadia fit in? As in, will Nadia’s hardware design and capabilities make developing software to get her to do stuff any easier?
That’s a really good question. We definitely have done a lot of things that are work-arounds for less “capable” robots. For example, you don’t go up a stair step the same way really slowly that you would at normal speeds. We’ve also done a lot of controls focus for balancing with slower robots. We think that a lot of this will translate really well to a fast platform, but know that it will require some philosophy changes.
For example, when walking slowly with our current Atlas, we don’t do a whole lot of step adjustment, because it isn’t that helpful for balance. However, when walking quickly, that’s likely to be our main balance strategy. We do think that things will be just easier to get working on a faster robot, though. There has been a lot of cool work on open-loop gait trajectories, and pretty much all of them are for (really) fast-moving robots. I think this embodies that the higher speeds are just more stable and forgiving than slower ones.
You say that Nadia will have “size, weight, speed, torque, and range of motion comparable to that of a typical human.” In this context, can you talk about the design philosophy and capabilities of Nadia relative to a robot like Kengoro that’s designed to mimic a human as closely as possible?
Unlike Kengoro, we’re not going to use a ton of redundant actuation. Humans are incredible in that, regardless of our configuration, we always have some subset of muscle that can be used, avoiding singularities throughout our ranges of motion. This seems like the strategy that Kengoro is going for, using these redundant antagonistic actuators.
For actuation, we’re going for more traditional linear hydraulic actuators with a few rotary electrics thrown in where it makes sense. This makes it really tricky to design the mechanisms, though, because we have to avoid singularities and maintain some kind of adequate mechanical advantage at all times. We’ve used a lot of simulation data to get our benchmark points, and then use simulation and 3D printing to determine whether or not these mechanisms will be sufficient. This has worked really well, too, because we’ve had to discard a few designs that looked really promising, but then had issues like too much flexibility when mocked up.
The models of Nadia don’t appear to show any onboard power or hydraulic pumps. How different will the robot look from the 3D-printed models?
The final robot will look significantly different from the models. The model in the pictures is more meant just to prototype mechanisms and see how they will work, and isn’t really representative of the structure (endoskeleton/exoskeleton). Also the real robot isn’t likely to be near as colorful. We’re targeting the height and weight of a human, as well as being in the ballpark of human volume. So design targets are between 5’7″ and 6’0″, and sub-90 kg.
What’s the plan for Nadia’s head? And will Nadia be able to operate autonomously, or will it be more of an “avatar” with a human directly in the loop somewhere?
We’re looking at a couple of different head designs, but are kind of holding off on that for now, focusing more on the legs and torso, as the head doesn’t have much of an influence on the rest of the body. We’re big proponents of humans being somewhere in the loop, with that location being easily changeable. What I mean by that is we want the robot to be able to function autonomously, so that a human gives high-level behavioral instruction (e.g. go close this valve in this room), but have the human able to jump in at any part of the stack, from modifying a lower level part of the autonomy (e.g. that’s the wrong valve, I meant this one) to fully taking control of the robot (e.g. directly controlling hand poses to turn the valve).
What’s the development timeline for Nadia?
The project is a 3-year project, starting this last January. That’s really about the most I’m willing to commit to, as far as timeline. We’re really hesitant to say that when you’ll be see a Nadia walking around, because we want to design the robot properly and not rush things. But I can say that one of the final ONR project goals is to show Nadia performing tasks autonomously.
When completed, Nadia will have a greater range of motion than most humanoid robots, along with speed and power comparable to that of a human of the same size. Does this mean that a humanoid robot gymnast is possible? According to Jerry Pratt, who leads the robotics group at IHMC, “the answer is no, but we can move towards that goal.” Fortunately, Nadia (the robot) won’t need to match Nadia (the gymnast) in order to operate in complex human environments, and it’ll still be able to take the place of a real human in dangerous situations.
We’ve also heard that Nadia hardware will be developed and supported by Boardwalk Robotics, a so-new-it-doesn’t-even-have-a-website spinoff of IHMC, founded by Jerry Pratt. This kind of model seems to be getting more common in robotics—it helps research institutions develop products rather than one-off research robots, while also helping research-grade hardware make it out to other labs and then (hopefully) into industry.
This is only our very first, very early look at Nadia, of course. And as Griffin says, the robot is going to look much different from the mockups, and three years is a long time in robotics. We’ll be following IHMC’s progress on Nadia as closely as we can, and bring you updates as soon as we have them. Or sooner than that, if we can swing it.
[ IHMC ]