How NASA Is Using AI to Develop the Next Generation of Self-Driving Planetary Rovers
With each new generation of NASA’s Mars rovers, improvements in AI, machine learning and other advanced technologies continue to make them more autonomous as they traverse the Martian soil looking for clues about the history of Earth and our solar system.
But there is still more work to be done to give future planetary rovers even more powerful autonomous self-driving capabilities, according to Shreyansh Daftry, a robotics technologist with NASA’s Jet Propulsion Laboratory in Pasadena, California. Daftry presented a talk on the topic March 31 (Wednesday) at the virtual AI Systems Summit put on by Kisaco Research.
The latest Mars rover, Perseverance, landed safely on the red planet on Feb. 18, filled with new innovations and grandiose plans for exploring the Martian surface for the next two years. Perseverance already uses early forms of AI to find its way around Mars, but like the Curiosity rover before it, its autonomous systems work at speeds far slower than if humans were on board to drive them over the surface, said Daftry.
That’s where future AI, ML and other advancements are aimed at making even broader improvements as researchers and engineers look to developing better, faster rovers to work with humans as they explore planetary destinations beyond Earth's moon, he said.
“My team has been working on developing autonomy and guidance that could help our rovers drive just like a human driver would,” Daftry said. “While we understand that it would probably take hundreds of years for AI to match human intelligence, even if it could just model some of those behaviors it would help improve the driving capabilities of our rovers so they would be able to drive significantly safer, faster and further than we do today.”
To do that, much more data is required including the processing of huge numbers of images from the rover itself as it traverses the terrain, he said.
NASA’s earlier Curiosity rover, which landed in August of 2012, is guided on Mars each day by a team of human researchers. The researchers upload commands to the rover to direct it where to go and what to do to accomplish its mission the next day. Those procedures are necessary due to the time it takes for data and commands to be sent or received from Earth over the 293 million miles between the planets.
“This has been the typical way NASA has operated and developed missions in the past, and this has worked so far,” said Daftry.
With the new Perseverance rover, however, its algorithm has been improved to allow the latest rover to be operated continuously, without having to stop to crunch numbers.
Daftry wants to build upon that progress on future rovers using AI.
“Overall, our long term vision is to create something like Google Maps on Mars for our rovers so the human operator can specify the destination where they need to go, and the rover uses the software to find that spot,” he said.
Making Mars Rovers Faster
While autonomous rovers today can’t move as quickly as scientists would like, past space vehicles have shown it is possible, including NASA’s lunar roving vehicles, which accompanied three crewed Apollo missions in 1971 and 1972.
Those lunar rovers were driven by U.S. astronauts across the moon’s surface at comparatively high speeds, almost 100 times faster than today’s rovers can move across Mars, said Daftry.
The key difference between how Perseverance moves today on Mars and how the lunar rovers moved across the moon in the 1970s is that the lunar vehicles were powered by humans, without the need for AI, he said.
And even though NASA has built many advanced exploratory vehicles and machines in the past, “they're not really what I as a trained roboticist would yet call ‘intelligent,’” said Daftry.
The current operating procedures for Perseverance and Curiosity are not scalable for future missions, especially if they also involve humans traveling to planetary destinations which are even farther from Earth or setting up colonies on Mars, he said.
“There's a growing need for future spacecraft to be autonomous, self-aware and have the ability to make critical decisions on their own,” despite the challenges of distance, said Daftry. “Imagine having a team of 100-plus engineers and scientists sitting here on Earth, controlling each of those assets. Sounds crazy, right? The only way that we can scale up the space economy is if we can make our space assets self-sustainable. And artificial intelligence is going to be a key ingredient in making that happen.”
An extension of that could be the possibilities of sending human astronauts on future missions along with robots, which would create human-robot teams that could work together, said Daftry.
“A lot of the work right now is under technology development and our hope is to infuse them into our future Mars rovers in the next three to five years,” he told EnterpriseAI. “Some may be easier to integrate already but others may need more maturity to ensure it doesn’t compromise safety. We are still far from solving the problem of trustworthy AI systems.”
First versions of these coming systems may include offering “suggestions” to the next-gen systems to aid in decision making, and maybe eventually using “black box” decision-making once verification and validation steps can be incorporated, he said.
“In some ways we have been already using AI on Mars for decades since it runs computer vision algorithms for autonomous localization” on rovers, he said. “The Mars Exploration Rover and the Curiosity rover have been doing fully-autonomous navigation drives on Mars for more than 10 years. We even have a computer vision-based detection system (AEGIS) that runs onboard the Perseverance rover and finds interesting rocks to sample.”
These past projects have pretty much all used classical AI and robotics model-based techniques, though, said Daftry. “If we define AI in the context of data-driven approaches, we are only getting started, and this is where my team and I are pushing to infuse ML/DL and data-driven approaches in general. Some technologies are easier to infuse like semantic segmentation because you can easily put safety checks on top of it. But other aspects like onboard adaptation, where you let your AI system learn in-situ, is a totally different regime because it breaks the very fundamentals of how we do system verification, and needs us to rethink how to build trustworthy systems.”
Progress is being made on these projects and Daftry says his dream is to deploy the first deep learning algorithm on a Mars rover in the next three years, with extensive AI-based, on-board autonomy as a central piece for all rovers in the next 10 years.
“I see this as one of the greatest grand challenges of AI and robotics in this century,” said Daftry. “One of the things that's very close to my heart is human exploration. I would love to work towards how these robots can become even more intelligent so that they can really work with human teams in a very seamless fashion. It has a huge potential to revolutionize space exploration in the future, but there are a lot of key challenges that still remain.”