Fed Gov Today Teaching Space Robots to Think: Naval Research Lab’s Reinforcement Learning Breakthrough
Sep 16, 2025
Samantha Chapin, a space roboticist at the Naval Research Laboratory, discusses groundbreaking work in teaching robots to think autonomously using reinforcement learning, akin to dog training. She shares insights on how these robots adapt to unpredictable environments, particularly aboard the International Space Station. The conversation highlights the importance of simulation and open-source tools from NASA, paving the way for future advancements in autonomous space assembly, satellite servicing, and reducing the need for constant human oversight in space missions.
AI Snips
Chapters
Transcript
Episode notes
Learning By Reward
- Reinforcement learning trains agents using rewards from environment interactions, similar to dog training.
- Kenneth Stewart explains this lets robots learn behaviors by trial and reward rather than fixed commands.
Simulation For Robustness
- Training in simulation with varied parameters creates policies robust to real-world mismatch and unexpected payloads.
- Samantha Chapin says this domain randomization helps robots perform even if reality differs slightly from models.
Astrobee Enabled First Space Test
- NASA's existing Astrobee on the ISS enabled the team to test complex autonomy without launching new hardware.
- Samantha Chapin recounts the policy worked in the five-minute ISS test on the first try, marking a key milestone.