Every chance we get, we post videos highlighting the adventures of MARLO, the University of Michigan’s blissfully unaware bipedal robot. MARLO is totally “blind,” without cameras, lidar, or anything else to show it where it’s going. But the robot is still able to walk dynamically over a range of terrain that I think would be appropriate to call staggering. Varied terrain does indeed stagger MARLO on a regular basis, and it’s probably fallen over more times on video than any robot we’ve ever seen.
Professor Jessy Grizzle and his students have been challenging MARLO with increasingly difficult terrain, most recently at a location on the beautiful Ann Arbor campus called the “Wave Field,” an “earth sculpture” created by artist Maya Lin. A new video posted today shows off MARLO’s latest exploits, and we talked with Grizzle about why a robot that falls over all the time is the key to efficient dynamic humanoid walking.
MARLO doesn’t look like most other humanoid robots, and it’s not just the pronounced lack of arms. Or a head. It’s underactuated, which means that its legs have more degrees of freedom than motors to control those degrees of freedom. MARLO’s ankles don’t have actuators in them: they’re “floppy,” to put it technically. “This type of locomotion is more agile,” Grizzle tells us, “but also requires more sophisticated control algorithms than a fully actuated robot such as ATLAS and ASIMO, which almost always execute a flat-footed walking gait.”
The curious mechanical design (using a four-bar linkage for the legs) allows MARLO’s motors to be placed up where its hips are, keeping the legs free of actuators, gear boxes, and cables. As a result, MARLO has some of the thinnest and lightest legs you’ll ever see, which leads to more durability and hardware that’s easier to fix, very important when you have a robot with a penchant for tumbles.
Hardware is only one part of the magic required to get a robot like MARLO to walk reliably. Grizzle says his team tries to discover “fundamental mathematical principles” for highly dynamic bipedal locomotion. “We want our algorithms to work for whole classes of bipeds and not just our robot,” he explains. “Whatever we do, we try write it up in clear enough terms so that others can implement our results without calling us up and asking ‘what’s the secret sauce?’ ” He continues:
“We use the full dynamical model when designing our controllers, and not a simplified inverted pendulum model or a spring-loaded inverted pendulum model. The field has accomplished a lot with the simplified models, but they have limits: when people use simplified models, they end up with slow, flat-footed walking. We use optimization algorithms on our model to design gaits that respect torque and speed limits our motors, friction cone constraints on ground contact, and take into account during the gait design a range of conditions the robot will encounter, such as changes in ground height, lateral slopes and walking speeds.”
With skills like these, MARLO has no problems tackling all kinds of rugged terrain. Almost no problems. Well, maybe just a few problems. But these problems are really more like teachable moments, and they only come up when MARLO is being made to do some crazy stuff. Here it is walking on the Wave Field, which Grizzle subjected it to “because it is there.”
You can watch some behind-the-scenes footage here, part of a late-night indoor practice session that makes me think MARLO may be my upstairs neighbor.
Grizzle says that when it comes to MARLO trying to conquer the “devious undulations” of the Wave Field, “we have gotten farther than I thought we would, to be honest.” This is the sort of thing we like to hear from researchers and don’t, usually: pleasant surprise about how well their robot is performing. For more details on how MARLO managed to get this far, we asked him a few specific questions.
IEEE Spectrum: How does stepping continuously help MARLO keep its balance?
Jessy Grizzle: In the videos, MARLO is swapping from one foot to the other between two and three times per second. Why? When MARLO is not moving forward, stepping in place is necessary because the ankles in the feet are pivots. If she stood still, she’d fall forward or backwards: imagine standing on a rotating bar. By cycling the legs, they can be placed slightly in front of the center of mass or slightly behind, to maintain balance, the same as rolling the wheel of a Segway back and forth to keep the rider (an inverted pendulum) upright. Because a wheel is continuous, you do not even notice the small forward and backward motions. With legs, you see the up and down motions, even if they are almost being put back down in the original positions, because the corrections are very small.
When MARLO falls, what’s usually the cause? Do you learn more from the falling, or the walking?
Failure almost always provides more information than success, but if you never had success, it would be totally depressing!
When we are moving faster than 0.5 m/s, falls mostly occur from the foot slipping on the ground, or from tripping over an object that was higher than we “told” the robot to expect (she is blind). We “tell” the robot what to expect by including terrain information in our gait design algorithms. For example, we impose that the optimizer find, if feasible for the robot, a periodic walking motion that can withstand step-to-step ground height variations of +/- 5 cm.
In our previous Wave Field video, we fell because the lateral leg motion did not take into account the steep slopes of the ridges and knolls running down the Wave Field. The robot would seek to place the leg 15 cm to the side for example, but bang into the slope after only 5 cm, and fall down. In the current controller, the PhD student Xingye (Dennis) Da is estimating local terrain slope from the relative height changes of the feet at each foot placement. This information allows a trajectory to be planned that allows the foot to be moved where it is needed without banging into the side slopes.
The last fall in the video is for the reason explained in the video, and one reason why we do need a camera: the robot successfully climbed to the top of the knoll. The foot falls then took place at roughly even height, so the ground must be relatively flat. The robot had no information that she was about to walk off a cliff. She took one step and never reached bottom before titling over to the point that our software automatically killed the power.
With MARLO doing such a good job walking without a camera, why is it important to incorporate vision next? What kinds of things will that enable?
Walking blind does not allow the robot to adapt to severe terrain changes. For example, we could not walk on flat ground and then randomly place a set of stairs in front of MARLO and expect a felicitous outcome! With a camera, we could easily switch gaits and commence climbing… at least, we claim we can. We have not done it.
Something we wish to attempt is the obstacle course of the W-Prize. We will not attempt the energy portion of the prize, but the obstacle course seems to be completely beyond anything attempted by a bipedal robot to date. We believe it is doable; not with MARLO, but with the new Cassie-series robot [pictured below].
Is there a timeline for Cassie, and where will vision fit in on either MARLO or Cassie?
My hope is to purchase robot No. 1 off the Cassie production line! Jonathan [Hurst] has added a lot of spice to my career through his wild robot designs. They push us control designers to the bleeding edge of sanity. I expect Cassie to be a very challenging, hair-pulling nightmare of a robot that will drive my team and me crazy, but we will have a blast figuring it out. I hope to place an order in October 2016 and take delivery in February 2017.
In the meantime, do not write off MARLO. My team is wickedly creative. We will keep developing ever better mathematics for walking, higher performing control design algorithms, and we’ll find out just what are the real limits of an ATRIAS series robot. Our math allows us to take into account the physical limits of the motors and the ground contact forces with the environment, and with Aaron Ames’s super fast optimization algorithms, we’ll design some gaits that will surprise everyone. And besides, the Wave Field is not yet conquered by any means.