← Back ICRA 2026

Learning Traversability Cost Maps with Decomposed Uncertainties Via Continuous-State MEDIRL

Gwanhyeong Song, Dongjae Lee, Ayoung Kim

PDF

AI summary

Key figure (auto-extracted from paper)

A modified MEDIRL framework successfully disentangles aleatoric and epistemic uncertainties in traversability cost maps, enabling accurate risk assessment and expert trajectory replication.

traversability cost maps uncertainty decomposition MEDIRL continuous-state MDP inverse reinforcement learning risk-aware navigation

Problem

Existing traversability cost maps lack reliable uncertainty quantification and kinematic fidelity, hindering robust risk management and safe autonomous navigation.

Approach

The authors extend MEDIRL by integrating a CVAE and decoder ensemble to separate data noise from model ignorance, while using continuous-state rollouts and a margin loss for stable, kinematically realistic training.

Key results

Disentangles aleatoric uncertainty for path ambiguity and epistemic uncertainty for unobserved terrain
Expected state visitation frequencies closely match expert demonstration trajectories
Accurately assigns high costs to physical obstacles like curbs and trees
Margin loss ensures stable convergence by penalizing low-cost generated rollouts

Why it matters

Provides a crucial foundation for safer autonomous navigation by enabling context-aware risk management through reliable, uncertainty-aware cost maps.

Abstract

Accurate traversability assessment is critical for mobile robot motion planning, yet sensor occlusions and model limitations often compromise cost map reliability. Therefore, analyzing spatial uncertainty is essential for robust risk man- agement. We propose a novel Maximum Entropy Deep Inverse Reinforcement Learning (MEDIRL) framework that learns a traversability cost map while explicitly disentangling aleatoric and epistemic uncertainties. Aleatoric uncertainty is captured via latent sampling in a Conditional Variational Autoencoder, while epistemic uncertainty is estimated using a decoder ensem- ble. For kinematic fidelity, we introduce efficient continuous- state rollouts utilizing precomputed transition grids and bilinear interpolation. Fusing camera and LiDAR features, our model achieves stable convergence guided by a novel margin loss. Re- sults demonstrate that learned state visitation frequencies match expert trajectories, and the decomposed uncertainties effectively identify high-risk terrains, providing a crucial foundation for safer autonomous navigation.

Index terms

Reinforcement Learning Imitation Learning Field Robots