CUTE-Planner: Confidence-aware Uneven Terrain Exploration Planner

TL;DR

The paper tackles autonomous planetary rover navigation in uneven terrain under elevation uncertainty by integrating Kalman-based DEM estimation with confidence mapping into a graph-based exploration framework. It introduces traversability-based sampling, adaptive online confidence updates, and confidence-constrained exploration to actively reduce map uncertainty in low-confidence regions while maintaining safety. A novel low-confidence region ratio metric evaluates map reliability, and extensive lunar-terrain simulations show substantial improvements, including up to a $69\%$ reduction in uncertain terrain and a $100\%$ mission success rate. The approach enhances both immediate navigation safety and long-term map quality, with implications for robust autonomous exploration in challenging planetary environments.

Abstract

Planetary exploration robots must navigate uneven terrain while building reliable maps for space missions. However, most existing methods incorporate traversability constraints but may not handle high uncertainty in elevation estimates near complex features like craters, do not consider exploration strategies for uncertainty reduction, and typically fail to address how elevation uncertainty affects navigation safety and map quality. To address the problems, we propose a framework integrating safe path generation, adaptive confidence updates, and confidence-aware exploration strategies. Using Kalman-based elevation estimation, our approach generates terrain traversability and confidence scores, then incorporates them into Graph-Based exploration Planner (GBP) to prioritize exploration of traversable low-confidence regions. We evaluate our framework through simulated lunar experiments using a novel low-confidence region ratio metric, achieving 69% uncertainty reduction compared to baseline GBP. In terms of mission success rate, our method achieves 100% while baseline GBP achieves 0%, demonstrating improvements in exploration safety and map reliability.

Figures (5)

• Figure 1: Process of the proposed confidence-aware uneven terrain exploration planner. (a) Traversability-based sampling: candidate vertices (orange circles) are sampled only in traversable regions and connected by orange lines to form the traversable local graph in the local traversability map. Blue circles and lines indicate the previous best exploration path. Lower traversability costs indicate terrain that is easier for robots to navigate. (b) Confidence-driven path evaluation: the planner selects a path that targets low-confidence regions (gray box) in the local confidence map; the green path represents the confidence-driven exploration path. In the confidence map, lower confidence scores represent higher uncertainty in elevation estimates. (c) Autonomous exploration path: leveraging both traversability and confidence information produces the best exploration path (pink), which balances information, safety, and map reliability.
• Figure 2: Confidence-aware autonomous exploration framework. A. Kalman estimator for DEM construction and confidence mapping, B. Traversability analysis from terrain attributes (slope, roughness, step height), C. Autonomous exploration using traversability cost and confidence score.
• Figure 3: Lunar simulation environments: (a) Moon 1, (b) Moon 2.
• Figure 4: Average explored volume over time in Moon 1 and Moon 2.
• Figure 5: Qualitative comparison of exploration results in Moon 1. Orange graph shows the planning graph generated during exploration. (a) GBP, (b) Only_Trav, (c) Ours.