Effectiveness and safety of PPL on real robots

Determine the effectiveness and safety of Predictive Preference Learning from Human Interventions (PPL) when deployed on real robots operating in physical environments, by evaluating its performance and safety characteristics outside simulation.

Background

The paper introduces Predictive Preference Learning from Human Interventions (PPL), an interactive imitation learning algorithm that forecasts future trajectories and converts human interventions into contrastive preference labels. This approach aims to reduce human effort and improve sample efficiency by propagating corrective intent into predicted, potentially risky future states.

All empirical results reported in the paper are obtained in simulation environments, including MetaDrive for autonomous driving and Robosuite for robotic manipulation. Consequently, the authors explicitly note that transferring PPL from simulation to real-world robotic platforms has not yet been evaluated and remains an open issue, specifically regarding effectiveness and safety.