Earlier this month, MIT posted a video of Mini Cheetah, a small quadruped robot from the lab of Sangbae Kim. We wanted to make sure you saw the video as soon as possible, which is why we featured it in Video Friday, but of course we wanted to know more about the robot. So last week we spoke a little bit with Sangbae and asked him a few questions about Mini Cheetah’s design and his future plans for the robot.
For some context about this robot, MIT notes that Mini Cheetah weighs about 20 pounds, with 12 modular motors that give each leg a powered hip joint (with 2 degrees of freedom) plus a knee joint. The motors (now off-the-shelf and cheap rather than custom made and expensive) have a single stage 6:1 gearbox, and we’re told that they have a maximum joint speed of 40 radians per second, which is (in technical terms) frickin’ fast.
The backflip you see in the video was, somehow, accomplished on the first attempt, and for its next trick, Mini Cheetah is being taught to perform the very cat-like maneuver of landing on its feet after being thrown into the air.
Sangbae Kim: This is mostly for research acceleration. If you talk to anyone working with robots they can tell you how painful it is: Robots are fragile, dangerous, not enough torque, hard to model. There are a bunch of problems. If you have a robot like Cheetah 3, what is the actual time that you’re running it? It’s probably 1 percent or less, because it’s hard to run, it’s dangerous, or if something breaks, it’s down for a month. With Mini Cheetah, you can run it like 5 hours a day. We’ve done 3-hour long demos with no problem. Basically, we’ve reached a level where it’s so reliable that we can give it to other groups that are suffering a lot with their hardware.
Mini Cheetah is just about the perfect size. Twenty pounds (9 kilograms) is not too small but not so big that it’s dangerous or fragile. This robot—we used it for 12 months, and we didn’t have to change a single mechanical component. It’s mainly our actuator technology. We designed the machine to be able to absorb the impacts, jumping and landing and so on. We’ve reached 2.5 meters per second, which is already fast for its size, but in theory it can run up to 4 meters per second—that’s reaching the maximum speed of Cheetah 3.
Hydraulic motors are still being used in most high performance dynamic robots. Why are electric actuators best for robots like Cheetah and Mini Cheetah?
I had a discussion about this with Marc Raibert—he and I have slightly different opinions. He thinks there are some scaling issues, but I disagree. Electric motors scale up really well, and I was honestly surprised at Mini Cheetah’s performance, because electric motors scale down badly. The secret is details in the design: If you look at the shoulder [actuators] of Mini Cheetah, they’re bigger, relatively, than in Cheetah 3. It’s not very noticeable, but Mini Cheetah has shoulders that are bigger relative to its body length, making it more powerful.
Electric motors scale up so easily that making a horse-sized robot, you’d be able to make it run as fast as a horse. There’s no issue in power density. I think the question becomes interesting when it comes to humanoids, because humanoids require so many degrees of freedom. So up to 12 degrees of freedom, that’s no problem, and I think electric motors are the way to go. But if you have 24 degrees of freedom, and you need to have multiple degrees of freedom in the wrist and ankle and neck… there are so many degrees of freedom all jumbled up in one ball and socket joint, that’s where electric motors tend to suffer. That’s where hydraulics starts becoming attractive, because you have this one big power source, but the actuators are relatively small, and a lot smaller than an electric motor with equivalent torque. There are other prices to pay for hydraulics, of course, so I think up to a certain number of degrees of freedom, I think electric actuators would be hands-down better in every aspect. But once the degrees of freedom get to be too much, hydraulics has a huge advantage.
Now that you have Mini Cheetah, are you going to continue working with Cheetah 3?
So far we’ve been working 95 percent in simulation, and then just testing on the robot when we have confidence in our controller. Now we’re going to move to something like 50 percent in simulation, 40 percent in Mini Cheetah, and 10 percent in Cheetah 3. Even if your algorithm in simulation isn’t super robust, Mini Cheetah has no problem with testing it. It takes like 2 minutes to set up, and if something goes wrong, it won’t break.
Not breaking is a big deal for research robots, because it means that you can try things on them, or trust your students to try things on them, without first setting up quadruplicate e-stops, putting on body armor, getting out your checkbook, and standing next to the fire alarm. As Sangbae points out, a rugged robot means that you can work less in simulation, which is much more fun (and potentially more productive) and seriously how else do you come up with all of those outtake videos?
The plan is to build 10 Mini Cheetahs, five or six of which will be pressed into service in different MIT labs to tackle different problems all at once, hopefully accelerating development by a factor of, well, five or six, I suppose. The other four or five robot could make their way out of MIT to groups that have experience with legged robots. Eventually, the hope is to have, according to Sangbae, “a robotic dog race through an obstacle course, where each team controls a Mini Cheetah with different algorithms, and we can see which strategy is more effective.” The race will likely be human-controlled initially, but eventually, Mini Cheetah will be going fully autonomous. And once it does, we’re keeping our fingers crossed for an affordable consumer version in the near future.
[ MIT ]