Artificial intelligence could be a huge help to Mars rovers like NASA’s Curiosity or Perseverance, but first these A.I. systems need to be trained on what to look for. A NASA project invites members of the public to help identify features of the Martian landscape, in order to train an algorithm that future rovers could use to navigate around the red planet.
The AI4Mars project was launched last year, and users have already labeled nearly half a million images to help develop the Soil Property and Object Classification (SPOC) algorithm. This algorithm identifies features of the landscape like sand and rock, and does so correctly nearly 98% of the time. In the future, this algorithm could be incorporated into Mars rovers’ autonomous driving capabilities like the AutoNav technology used by Perseverance.
Now, the researchers want to expand SPOC to get more detailed information about rock formations such as the presence of float rocks or nodules. By automatically classifying the types of rock imaged by rovers, the researchers can send driving instructions back to the rovers more quickly.
“It’s not possible for any one scientist to look at all the downlinked images with scrutiny in such a short amount of time, every single day,” explained Vivian Sun, a JPL scientist who helps coordinate Perseverance’s daily operations and consulted on the AI4Mars project. “It would save us time if there was an algorithm that could say, ‘I think I saw rock veins or nodules over here,’ and then the science team can look at those areas with more detail.”
To help develop this algorithm, NASA is inviting members of the public to go to the AI4Mars page on Zooniverse and look at images of the Martian surface captured by the Curiosity rover. They are asked to draw polygons around particular features like sand, soil, bedrock, and large rocks. The results of thousands of classifications made by the public are then collated and checked by scientists to make sure that the labeling is accurate.
Over time, as more individual pieces of data are labeled, the algorithm can learn to distinguish features for itself.
“Machine learning is very different from normal software,” said lead researcher for the AI4Mars project, Hiro Ono. “This isn’t like making something from scratch. Think of it as starting with a new brain. More of the effort here is getting a good dataset to teach that brain and massaging the data so it will be better learned.”
