Engineering students at ETH Zürich used simulation, AI and 3D printing to design this unique low gravity explorer.
Although humans may be returning to the moon by the end of this decade, robots are set to remain the workhorses of solar system exploration, as they have been for decades. Engineers are trying to figure out how to make them better at the task, and artificial intelligence (AI) is proving an indispensable assistant.
The problem with solar system bodies is that their masses, and therefore gravities, differ from Earth’s. It’s not such a big deal for planets like Mars, which, with one third of Earth’s gravity, can still comfortably hold onto objects on its surface. But on smaller bodies such as moons and asteroids, mechanical systems don’t behave as they do on Earth, said Fabio Bühler, a robotics research assistant at the ETH Zürich university in Switzerland.
“Classical rover systems that use wheels, like the Mars rovers, really struggle in low gravity environments,” Bühler, who led a team designing an experimental robot that could help explore small solar system bodies, told Engineering.com. “There’s no traction [in low gravity], so you can’t really drive in these environments.”
Even Earth’s moon, with a gravity six times weaker than that of Earth, can pose a challenge for wheeled vehicles, especially on slopes or in rough terrain. Bühler and his team of students decided to build a hopping robot, called SpaceHopper, that would be stable, independent and lightweight enough to explore distant mini worlds on its own. Here’s how they designed this fascinating style of locomotion.
Jump like a cat
Instead of wheels, SpaceHopper’s body rests on three legs with bendable joints, and it uses motors located in its hips and knees to propel itself upward. The student designers settled on three legs to give the robot stability when it rests on the surface, but also to keep its mass within limits as would be required for a future space trip.
“We designed this robot as a technology demonstrator to show that it’s possible to do that with current state-of-the-art technology legged robots,” Bühler said. “We envision that this kind of robotic explorer could move around low gravity bodies, for example to search for rare minerals and look for suitable locations ahead of a bigger mission.”
The students used Siemens NX CAD and simulation tools to rapidly iterate the design of the robot’s mechanical components. The electronics were designed with the open source KiCAD EDA platform. The main challenge, Bühler said, was to make the parts as lightweight and as compact as possible. The ETH Zürich team then 3D-printed the most promising solutions and tested them in the real world.
Nvidia’s Isaac simulation software subsequently helped the team to optimize the design of the robot’s legs and their locomotion. Taking inspiration from the ability of cats to always land on their feet when dropped, the students wanted the robot to be able to twist itself mid-air using just the inertia of its limbs to assume a correct position for a safe impact ahead of the next jump.
“The robot needs to land in a predefined orientation, otherwise, it’s not going to be able to use its feet,” Bühler said. “Traditionally, space engineers would use flywheels to stabilize the orientation of a robot, but that’s going to be an additional subsystem that needs more engineering and adds additional weight. So, we decided to use the robot’s legs to do this, the way cats do.”
2,000 twins
In their simulation, the students created a digital asteroid surface with programmed microgravity conditions similar to those on the asteroid Ceres. Some 950 kilometers wide, Ceres is the largest rock in the asteroid belt between Mars and Jupiter where the majority of solar system’s asteroids reside. Due to its size, Ceres was the first space rock to have been discovered by astronomers in the early 1800s. Nearly perfectly spherical, the asteroid was once considered a planet candidate and is still an object of intense scientific interest. Ceres was studied in great detail by NASA’s Dawn mission, which visited it in 2016 and found it was covered in ice and likely possessed a subsurface ocean. (The presence of water on the space rock led scientists to think it might harbor microscopic life forms.) But with a gravity of only 3% that of Earth, Ceres could pose a challenge for landers.
“The simulations are crucial because there’s no way for us to test it here on Earth and test it cheaply,” said Bühler. “So we use a physics simulator where we can simulate low gravity environments and try, for example, different leg lengths. We could vary the length of the shin and the thigh and look how these design choices perform in a simulation.”
After optimizing the robot’s design and demonstrating it could walk and jump on its three legs and land reliably on its feet, the engineering students used the same simulator to train the neural network that would be responsible for controlling the robot’s motion in the real world. They used a method known as reinforcement learning, which allows a system to figure out the most optimal behavior through its own unsupervised interactions with the environment in a trial-and-error fashion.
To speed up the training process, the students created 2,000 digital twins of the robot in the simulation. They instructed the twins to move independently around the digital asteroid Ceres, trying out things and learning in parallel from their mistakes. It took 12 hours to complete the basic training using an Nvidia GeForce RTX 3080 GPU.
“We used the proximal policy optimization algorithm to train the neural network,” said Bühler. “It’s really beneficial for us because we just tell the algorithm, for example, to jump in the most energy efficient way and then let it optimize the locomotion for energy efficiency.”
During the training process, the researchers kept altering some parameters of the robot’s system and mechanical behavior to teach the neural network to be somewhat flexible. Otherwise, Bühler said, the robot could struggle in the real world as no simulation is perfectly accurate.
“During the simulation, we need to vary some parameters, such as friction in the joints or the masses of the different limbs,” Bühler said. “That means the neural network doesn’t get really comfortable with the simulation environment but can generalize to other environments.”
Next stop — the moon
The 6-kilogram SpaceHopper is made of space-grade aluminum but features polymer-based 3D-printed parts inside its triangular body. The robot’s limbs are powered by motors made by Swiss firm Maxon, the same firm that produced electrical motors for NASA’s Perseverance Mars rover, which is currently exploring the red planet.
“The goal of the project was to make the robot as space-ready as possible,” said Bühler. “The mechanical parts are mostly space-ready. The electronics are not, but the motors that are inside the robot, for example, could be easily replaced one-to-one by space-ready models.”
The neural network trained in the simulator runs on an Nvidia Jetson Nano microcomputer that would have to be swapped for a space-hardened processor for a real space mission.
The team recently tested the robot in a parabolic flight simulating weightlessness to verify that the “cat” approach to reorienting its body mid-flight works as expected. For the flight, the engineers had to replace most of the robot’s polymer-based 3D-printed components with sturdier aluminum parts to make sure they survived the experiments.
“The 3D-printed parts are quite brittle and if it broke during the flight, there would be a lot of small plastics floating around the plane,” said Bühler.
The team that built the SpaceHopper has completed their degrees, Bühler said, but ETH Zürich will keep developing the three-legged robot concept hoping to adapt it for moon exploration. The next-generation hopping robot could make it to space for real.
