# How Does It Feel? Self-Supervised Costmap Learning for Off-Road Vehicle Traversability

**Authors:** Mateo Guaman Castro, Samuel Triest, Wenshan Wang, Jason M. Gregory, Felix Sanchez, John G. Rogers III, Sebastian Scherer  
**arXiv:** [2209.10788](https://arxiv.org/abs/2209.10788)  
**Date:** 2022-09-22  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2209.10788v1/

## Key points

- Method for estimating terrain traversability using self-supervised costmap learning
- Challenges of accurate physics models and informative labels in supervised learning for robot-terrain interactions
- Combination of exteroceptive environmental information and proprioceptive terrain interaction feedback in predicting traversability costmaps
- Incorporation of robot velocity in the costmap prediction pipeline
- Validation through navigation tasks on autonomous all-terrain vehicle (ATV) and large ground robot
- Smoother navigation and fine-grained understanding of interactions with different terrain types
- Comparison to occupancy-based navigation baseline on challenging off-road courses, reducing interventions by up to 57%
- Transferability demonstrated through short-range navigation experiment with another robot (Warthog)
- Innovative solution for estimating terrain traversability in off-road environments
- Effectiveness in improving navigation performance and reducing interventions on challenging terrains
- Transferability across different robotic platforms.

## Summary

The paper presents a method for estimating terrain traversability in off-road environments using self-supervised costmap learning. The authors highlight the challenges of building accurate physics models or creating informative labels for supervised learning in these complex interactions between robots and terrains. Their proposed method combines exteroceptive environmental information with proprioceptive terrain interaction feedback to predict traversability costmaps. They also introduce a novel way of incorporating robot velocity in the costmap prediction pipeline. To validate their method, the authors conducted multiple short and large-scale navigation tasks on a large autonomous all-terrain vehicle (ATV) and a separate large ground robot. The results show that using the learned costmaps leads to overall smoother navigation and provides the robot with a more fine-grained understanding of interactions with different terrain types such as grass and gravel. In terms of large-scale navigation trials, the authors compared their method to an occupancy-based navigation baseline on challenging off-road courses ranging from 400 m to 3150 m. They were able to reduce the number of interventions by up to 57% using their How Does It Feel (HDIF) navigation stack. This reduction in interventions demonstrates the effectiveness of their approach in improving navigational performance on difficult terrains with various features like slopes, vegetation, and tight turns. Additionally, the authors investigated the transferability of their method by conducting a short-range navigation experiment with another robot called Warthog. This analysis further confirms that their self-supervised costmap learning approach can be successfully applied to different robots. Overall, this paper presents an innovative solution for estimating terrain traversability in off-road environments through self-supervised costmap learning. The experimental results demonstrate its effectiveness in improving navigation performance and reducing interventions on challenging terrains, while also showcasing its transferability across different robotic platforms.