Adaptive-Critical: Identifying Critical States for Risk-Sensitive Navigation via Distributional Reinforcement Learning

Zhaofan Zhang1, Sihong Xie1,✉, Hui Xiong1,✉

1 The Hong Kong University of Science and Technology (Guangzhou)

✉ Corresponding authors

IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2026

Paper arXiv Code Video
Adaptive-Critical overview
Adaptive-Critical adapts its conservatism to state-dependent criticality signals estimated from a distributional value function, rather than relying on a fixed risk heuristic, producing safer trajectories under uncertainty.

Abstract

Adaptive-Critical is a state-dependent risk-sensitive framework that transcends fixed heuristics by integrating distributional criticality signals with sensory cues, enabling robust navigation under partial observability and environmental uncertainty. By estimating the full return distribution, the agent identifies decision-critical states where errors are most costly and modulates its risk preference accordingly, yielding safer trajectories without sacrificing efficiency in benign regions.

Overview

Placeholder for the method overview. Describe the distributional critic, the state-dependent criticality estimator, and how the risk preference is modulated during navigation.

Citation

@inproceedings{zhang2026adaptivecritical,
  title     = {Adaptive-Critical: Identifying Critical States for Risk-Sensitive Navigation via Distributional Reinforcement Learning},
  author    = {Zhang, Zhaofan and Xie, Sihong and Xiong, Hui},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026},
}