Enhancing Robot Navigation Policies with Task-Specific Uncertainty Managements

TL;DR

The paper tackles navigation under uncertainty where precision requirements vary by region and task. It introduces Task-Specific Uncertainty Maps (TSUMs) and a policy-conditioning framework called GUIDE to integrate region-dependent uncertainty into RL without reward engineering, formalized via the TSUM equation $U^\tau(l) = w_\Phi\,\Phi^\tau(l) + w_\mathcal{C}\,\mathcal{C}^\tau(l) + w_\mathcal{E}\,\mathcal{E}(l)$ and augmented state $\tilde{s} = [s,\, U^\tau(s),\, u(s)]$ within GUIDEd SAC (G-SAC). The method uses a RoBERTa-based parser to extract subtasks, spatial embeddings with triplet loss and attention to produce per-location uncertainty, and trains a Q-function and policy on the augmented state to focus uncertainty reduction where needed. Real-world ASV experiments show that G-SAC outperforms baselines across tasks, enabling safer navigation and more cost-effective use of precise localization by exploiting task-specific uncertainty requirements. This work provides a practical framework for task-aware uncertainty management in robotic navigation, reducing reliance on reward engineering and improving efficiency in uncertain environments.

Abstract

Robots navigating complex environments must manage uncertainty from sensor noise, environmental changes, and incomplete information, with different tasks requiring varying levels of precision in different areas. For example, precise localization may be crucial near obstacles but less critical in open spaces. We present GUIDE (Generalized Uncertainty Integration for Decision-Making and Execution), a framework that integrates these task-specific requirements into navigation policies via Task-Specific Uncertainty Maps (TSUMs). By assigning acceptable uncertainty levels to different locations, TSUMs enable robots to adapt uncertainty management based on context. When combined with reinforcement learning, GUIDE learns policies that balance task completion and uncertainty management without extensive reward engineering. Real-world tests show significant performance gains over methods lacking task-specific uncertainty awareness.

Figures (1)

• Figure 1: During pretraining, semantic and spatial embeddings are aligned via triplet loss and attention. At deployment, TSUMs derived from task descriptions and environment data condition the navigation policy for task-aware uncertainty management.