Reverse to Advance: Teleoperation-Cost Effective Hard Policy Learning from Reversed Easy Tasks

This paper introduces a method that leverages the directional asymmetry in task difficulty by using reversed easy-task trajectories as supervision for hard tasks. The authors propose a framework comprising an automated data collection pipeline, a hierarchical data refinement process involving temporal inversion and filtering, and an iterative policy learning loop. Experiments across two simulated benchmarks and real-robot setups demonstrate that this approach achieves higher success rates and better data efficiency for complex manipulation compared to reversal-based baselines and reinforcement learning methods, without requiring extensive manual teleoperation of the hard task.