Start Date
1-5-2026 12:00 PM
End Date
1-5-2026 1:00 PM
Description
Reinforcement learning enables robots to learn navigation policies in simulation, but performance often degrades during real-world deployment due to the sim-to-real gap. This work investigates domain randomization as a strategy to improve transfer for a Soft Actor-Critic navigation policy trained in a maze environment. Six targeted randomization methods are implemented including obstacle layouts, LiDAR noise, motion distance, positional noise, action perturbations, and maze switching to analyze how different sources of variability affect the robot’s ability to effectively navigate a maze. Results demonstrate the extent to which the robot can effectively generalize a policy for navigation given each domain randomization method, as well as which methods require adjusted reward shaping to prevent overfitting. This research contributes to the broader goal of helping to close the sim-to-real gap, with potential applications in exploration, mapping, and autonomous navigation.
Domain Randomization for Robot Navigation
Reinforcement learning enables robots to learn navigation policies in simulation, but performance often degrades during real-world deployment due to the sim-to-real gap. This work investigates domain randomization as a strategy to improve transfer for a Soft Actor-Critic navigation policy trained in a maze environment. Six targeted randomization methods are implemented including obstacle layouts, LiDAR noise, motion distance, positional noise, action perturbations, and maze switching to analyze how different sources of variability affect the robot’s ability to effectively navigate a maze. Results demonstrate the extent to which the robot can effectively generalize a policy for navigation given each domain randomization method, as well as which methods require adjusted reward shaping to prevent overfitting. This research contributes to the broader goal of helping to close the sim-to-real gap, with potential applications in exploration, mapping, and autonomous navigation.