# Safe Navigation in Unstructured Environments by Minimizing Uncertainty in Control and Perception

**Authors:** Junwon Seo, Jungwi Mun, Taekyung Kim  
**arXiv:** [2306.14601](https://arxiv.org/abs/2306.14601)  
**Date:** 2023-06-26  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2306.14601v1/

## Key points

- Presence of uncertainty in control and perception can lead to challenges like navigation failures and potential vehicle damage
- Novel framework introduced to minimize uncertainty in control and perception for safe navigation in unstructured environments
- Framework comprises a learning-based vehicle dynamics model and a self-supervised traversability estimation model
- Vehicle dynamics model quantifies epistemic uncertainty, enabling active exploration for efficient data collection and avoidance of uncertain state-action spaces
- Traversability cost prediction network uses meta-learning techniques to reduce aleatoric uncertainty by adapting online based on interaction experiences
- Integration with sampling-based model predictive controller optimizes trajectories to avoid uncertain terrains and state-action spaces
- Experimental results show significant reduction in prediction uncertainty and enhanced stability during autonomous navigation in unstructured environments
- Validation error comparisons demonstrate superior performance over baseline approaches across various terrain types
- Traversability cost prediction network effectively maps sparse LiDAR point clouds using voxel discretization techniques
- Framework offers promising solution for enhancing safety and reliability in autonomous vehicle navigation through minimizing control and perception uncertainty

## Summary

In the realm of autonomous vehicle navigation, the presence of uncertainty in both control and perception can lead to significant challenges, including navigation failures and potential damage to the vehicle. To address this issue, a novel framework has been introduced that aims to minimize uncertainty in control and perception to ensure safe and reliable navigation in unstructured environments. This framework comprises two key components: a learning-based vehicle dynamics model and a self-supervised traversability estimation model. The vehicle dynamics model is trained to quantify the epistemic uncertainty of the model, enabling active exploration that facilitates efficient data collection and effective avoidance of uncertain state-action spaces. Additionally, meta-learning techniques are employed to train a traversability cost prediction network capable of adapting online based on interaction experiences, thereby reducing aleatoric uncertainty. By integrating these models with a sampling-based model predictive controller, trajectories can be optimized to avoid uncertain terrains and state-action spaces. Experimental results have demonstrated that this approach significantly reduces uncertainty in prediction and enhances stability during autonomous vehicle navigation in unstructured environments. Furthermore, validation error comparisons for real-world driving data showcase the superior performance of the proposed method over baseline approaches across various terrain types such as unpaved surfaces, rough grasslands, profiled roads, and simulated environments. The traversability cost prediction network's pipeline overview illustrates its effectiveness in mapping sparse LiDAR point clouds using voxel discretization techniques. In conclusion, by leveraging advanced modeling techniques and adaptive learning strategies, the framework presented in this study offers a promising solution for enhancing safety and reliability in autonomous vehicle navigation through the minimization of control and perception uncertainty.