Last time we saw UC Berkeley’s Salto-1P, the robot was bouncing all over the place, able to hop continuously without any trouble. It was impressive to watch, especially when it occasionally exploded itself. But of course the Berkeley researchers had long-term plans to take the robot from generally bouncy to bouncy with a level of control that would allow them to teach it some sophisticated tricks. We saw some of those tricks this week at IROS 2018 in Spain, where Justin Yim presented a new video of Salto-1P turning some office furniture into a bouncy house.
Salto-1P’s hardware is the same as last year. It’s got one actuated elastic leg that gives it hops, an inertial tail that spins to control pitch, and half a quadcopter to control yaw and roll. The big difference this year is in the controller. The robot behaves more or less like a spring-loaded inverted pendulum, a simplified dynamic model that shows up often enough in both biology and robotics that it has its own acronym: SLIP. Way back in the 1980s, Marc Raibert developed a controller for SLIP-like robots, and people are still using it today, including Salto-1P up until just recently.
The reason why Salto-1P needed a better controller is because making the robot do things like you see in the video above requires very accurate foot placement—the robot has to be able to bounce on that chair without falling off, which means consistently landing more or less exactly where it wants to land after every jump. And even more aggressive behaviors, like jumping between walls, involves an entire sequence of jumps that all need to be accurate, since a single inaccurate jump will likely cause the robot to fall and smash itself to bits.
The fact that Salto-1P behaves ballistically means that it follows very predictable trajectories, which is good, but it also means that the only control that you have over the thing is during the excessively brief instant when it’s making contact with the ground. This can make things tricky, since most of the control that you get comes from the angle of Salto-1P’s leg at touchdown and the length of its leg (which affects how much power goes into the next jump), and that’s it.
In their tests, Salto-1P’s new controller proved to be a significant improvement over the original Raibert controller. Over a series of 95 randomish and occasionally aggressive hops, the standard deviation of Salto-1P’s foot placement error was just 0.1 meter, three times better that the error of a Raibert controller, and 95 percent of its touchdowns were within 0.3 meter of the targeted position. This is the kind of reliable, targeted jumping that allows the robot to hop on a chair, although anything much smaller than a chair is still a bit of a challenge at the moment.
As you might expect, the more aggressive Salto-1P’s jumping is, the less precise it is. Optimal precision comes after a series of small jumps, while the hyperaggressive long-distance flinging itself about that we’ve seen in the past is much more difficult to target. With a mid-to-low jump height, Salto-1P can handle both moving targets and surfaces that aren’t flat and level, like the office chair in the video. And it’s likely to get better at what it does, too: We spoke with Justin Yim at IROS, who told us that he’s working toward increasing Salto-1P’s jumping precision even more, while also weaning it off of the external localization and computing systems that keep it confined indoors.