FURI | Summer 2025

Bridging the Sim-to-Real Gap in Autonomous Systems Using Learning from Demonstration

This research examines the effectiveness of learning from demonstration (LfD) techniques in transferring knowledge from simulated to real-world environments for complex autonomous driving tasks involving obstacle avoidance. The study utilizes the Duckietown platform to compare LfD approaches given the data combination from simulation and real-world scenarios through metrics including average distance traveled and survival time. Results indicate that simulating data can aid the policy to achieve comparable performance while requiring fewer real-world interactions. This work provides insight into how to leverage task-relevant external data, facilitating effective knowledge transfer from simulation to real-world environments. Future research will explore combining LfD with other transfer techniques or conduct experiments with diverse driving tasks to better understand the effectiveness of this learning approach.