September 16, 2025
Subscribe and listen to the Fed Gov Today Podcast anytime on Apple Podcasts, Spotify, or at FedGovToday.com.
Human time in space is precious — and researchers at the Naval Research Laboratory (NRL) believe robots can help make the most of it. In a recent conversation on Fed Gov Today with Francis Rose, Dr. Samantha Chapin, space roboticist, and Dr. Kenneth Stewart, computer research scientist, explain how they are using reinforcement learning to teach space robots to work more autonomously.
Stewart kicks off the discussion by describing reinforcement learning in plain terms. “It’s basically where a robot or agent learns by getting rewards for doing the right thing,” he says. He compares it to dog training — you give a command, the dog obeys, and you reward it with a treat so it repeats the behavior. The same principle applies to robots: when they accomplish a desired task, they get a “reward” signal, and over time they learn to do that task reliably.
Chapin explains why this approach is so promising for space operations. Space missions are expensive and risky, and many current robotic operations are controlled manually — almost like using a joystick from Earth. “We want to make it so the robot is observing its environment and making decisions for itself,” she says. Reinforcement learning allows the team to train robots in simulation across thousands of slightly different scenarios. “We can vary things like the mass of the robot or unexpected payloads,” Chapin notes, “so that even if something isn’t exactly as expected, the robot can still perform the function we want.”
The researchers put this approach to the test using NASA’s Astrobee robot on the International Space Station. Stewart describes the team’s “crawl, walk, run” strategy, starting with simple movement commands. Using the Nvidia Omniverse simulator, they train tens of thousands of virtual robots in parallel within hours, test their algorithms on NASA’s open-source simulation environment, then move to hardware verification on a granite table before finally testing in orbit.
That method paid off: in just five minutes of real test time aboard the ISS, their algorithm worked the first time. Chapin emphasizes that the ISS is the perfect platform for this kind of work because it offers a real zero-gravity environment where results can’t be fully replicated on Earth.
Looking ahead, the NRL team sees applications for this technology far beyond simple navigation. Future goals include using autonomous robots for inspection, manipulation, and large-scale space assembly — potentially constructing massive structures like telescopes or power stations, or servicing satellites without sending astronauts. “You shouldn’t be afraid of this autonomy,” Chapin says. “We can make it robust and able to handle a bunch of different types of failures.”
Their next challenge? Taking what they’ve learned and testing it in the true vacuum of space, perhaps with a CubeSat mission. Each experiment brings them one step closer to a future where autonomous robots expand human reach in orbit and beyond.