EPFL’s Pleurobot is, obviously, our favorite robot salamander. This is likely because it looks so much like a real salamander, but more importantly, it moves just like a real salamander. Or, to be more specific, EPFL has spent years trying to make sure that the way Pleurobot moves is as close to the way that a real salamander moves as possible.
In a new paper just published in the Royal Society journal Interface, EPFL researchers describe how they’ve combined “high-speed cineradiography, optimization, dynamic scaling, three-dimensional printing, high-end servomotors, and a tailored dry-suit” to refine their robot to accurately capture the degrees of freedom, range of motion, and gait behaviors of the real animal.
Why so much focus on the humble salamander? Primarily, it’s because they’re cute, but there are a variety of much less important considerations that make salamanders interesting to study as well. They’re relatively primitive quadrupeds, making them simpler to study and model, but they also represent sort of a transitional animal between fish that swim and quadrupeds that walk, which (evolutionarily speaking) is an interesting place to be.
What’s even more fascinating is that the multimodal nature of the salamander—its ability to both swim in the water and walk on land—lies entirely in its musculoskeletal structure and nervous system. If you behead a salamander and then stimulate its spinal cord, what’s left of it will start to walk. Stimulate it more, and it’ll walk faster, and if you keep going, your headless zombie electrosalamander will transition into a swimming gait. Brains are overrated.
The reason that this is important, besides the obvious application of raising an army of the dead, is because robots (most robots, anyway) spend way too much of their brainpower trying to just keep from falling over. Shifting all of that to an intelligently designed structure with low-level control could be advantageous to walking robots, multimodal or otherwise. And of course, a robot that so closely replicates a biological system in biomechanics and neural control could be a fun thing for salamander-ologists to play with, but we’re a robotics blog, so let’s get back to that.
Pleurobot is what’s called an anchor model of a salamander, meaning that it’s “a realistic model fixed firmly or grounded in the morphology and physiology of an animal.” Doing this with a robot as opposed to in simulation is critical, because it’s simply not possible to accurately simulate the way that physical things interact with the world around them in software.
The design of the robot is based off of what sounds like a borderline insane amount of detailed study into the workings of real salamanders. We’re talking biplanar high-speed cineradiographic recordings of the animals walking, analyzed at 500 frames per second, and used to develop a robot scaled to just under 9x the size of the real thing. The size increase is due to Pleurobot needing DC motors, and efficient DC motors have a minimum size, so the robot was scaled around them (Dynamixel MX-64R servos, to be specific). Pleurobot uses 27 motors in total, with 11 spine segments, far less than an actual salamander. This was a somewhat arbitrary optimization, the researchers say, which still allows Pleurobot to “imitate the bending of the salamander’s spine in different gaits to a good extent,” according to the paper. The scaling was carefully calibrated to make sure that the robot’s movements, speeds, and forces exerted at its larger size are all comparable to the smaller living salamander.
Once you’ve got a good hardware model that you’ve tried to get to duplicate the motions of a salamander, you need to experimentally verify that your robot is, in fact, doing what you want it to do:
The aims of the experiments are (i) to demonstrate that we can reproduce the two basic behaviours of the salamander, so well that it has the potential to reproduce more complex locomotor behaviours if these are recorded from the animal and (ii) to validate whether replaying kinematics recorded from the animal, provided the scaling factors, can yield behaviours comparable to the salamander’s. Especially the second point will help us validating future experiments in which Pleurobot is driven by a neuronal model instead of pre-recorded animal kinematics.
The experiments themselves are straightforward: that x-ray movie of a salamander moving was compared with motion tracking of Pleurobot moving to see how closely they matched, and the results are pretty good:
There are a few differences: the stride length of the robot is a bit shorter than it should be, and the feet move a bit more quickly, but this is because Pleurobot has balls for feet and no toes. Swimming gaits are also quite similar, as long as they aren’t too fast, since at higher amplitudes, the motors can’t keep up. Overall, the scaled-up robot is able to replicate the animal’s gaits with remarkable similarity.
The fact that this works may not necessarily be surprising, considering that the Pleurobot is just playing back a sequence of postures and motions taken directly from the way that the animal moves. The researchers point out that with a good controller and actuators, getting a robot to move like an animal is indeed “trivial,” but what’s not trivial is getting it to keep moving in the real world while maintaining that similarity:
Unlike automatons in amusement parks that are solidly fixed to the ground, Pleurobot’s displacements in three-dimensional space are the results of (complex) physical interactions between the robot internal movements and the environment. The robot must be properly designed such that the interaction forces match those of the animal.
For instance, during ground locomotion, inaccurate mass distribution could lead to incorrect body orientations and incorrect contacts to the ground (e.g. a limb not touching the ground when it should) and hence different locomotor patterns and locomotion speeds. Similarly, the swimming of the robot could be very different from that of the animal if geometrical and dynamical properties of the robot had not been properly adjusted with the dynamic scaling. The indirect and direct comparisons between robot and animal presented here indicate that the interaction dynamics closely correspond.
So, great, EPFL has a very accurate robotic salamander, now what? First, being researchers, they want to do what researchers always want to do, and make it even more accurate. For example, since the experimental validation showed that toes are important, and that ball feet aren’t a good compromise, Pleurobot’s next upgrade will be all about “integrating as many aspects of the foot as possible given their complexity.” They also want better, back-drivable motors (or torque sensing motors with fast control loops) that can emulate the behavior of muscles, perhaps combined with elastic structural elements.
Once Pleurobot is perfect (or close enough that the roboticists are willing to stop fiddling with it for a little bit), there are lots of things that it’s useful for. Instead of driving it with pre-recorded animal kinematics, future experiments will try using neuronal models instead, which could make it into a useful tool for neuroscience research: “as a realistic physical model of the salamander capable of emulating basic and complex behaviours, it can serve to test hypotheses about the interactions between the different components underlying locomotion, in particular the interactions between descending modulation, central pattern generation, sensory inputs and interaction forces from the environment.”
In terms of what the robot is actually, you know, good for in a practical sense, that’s always a tricky question for research robots. An amphibious salamander robot capable of multimodal locomotion could hypothetically be used for search and rescue in disaster areas, but every researcher says that about every robot that’s able to move anywhere. We’re slightly more optimistic about Pleurobot for this kind of task (especially with that slick custom drysuit), but what’s perhaps more exciting is what we can learn about how to design robots that can use behaviors based on simple neuronal and musculoskeletal structures to take care of all of the tricky movement stuff, leaving overtaxed robot brains to ponder more important things.