DeRep: Learning Depth Representations for End-to-End Navigation in Cluttered Environments

This paper introduces DeRep, a navigation-oriented representation-learning method that integrates local bird's-eye-view reachability supervision with depth reconstruction. The authors pretrain encoders with a common architecture and freeze them during downstream learning to isolate the effects of representation objectives. Evaluation via linear policy probes, flow-matching imitation learning, and PPO shows that DeRep achieves the lowest prediction errors and outperforms reconstruction alone across all tested conditions. Both reachability-based objectives yield higher mean success rates in imitation learning than reconstruction alone. The method is deployed on a physical quadrotor in Gazebo without fine-tuning, demonstrating zero-shot sim-to-sim and sim-to-real transfer capabilities.