Learning Autonomy: Off-Road Navigation Enhanced by Human Input

TL;DR

The paper tackles off-road navigation by learning a local planner from short monocular-vision demonstrations. It builds a five-item utility vector $U(S,P) \in \mathbb{R}^{m\times5}$ that combines four terrain-based utilities and a distance-based utility, augmented by geometric cues from occupancy maps, and trains a neural classifier to map these utilities to trajectory choices from a fixed set $S$. By framing the learning as supervised classification over demonstration labels, the approach achieves rapid, data-efficient learning (approximately $5$–$10$ minutes) in simulation and demonstrates generalization to unseen terrain configurations. The method reduces manual tuning, avoids full dynamic modeling, and shows potential for sim-to-real transfer in diverse off-road environments with monocular input.

Abstract

In the area of autonomous driving, navigating off-road terrains presents a unique set of challenges, from unpredictable surfaces like grass and dirt to unexpected obstacles such as bushes and puddles. In this work, we present a novel learning-based local planner that addresses these challenges by directly capturing human driving nuances from real-world demonstrations using only a monocular camera. The key features of our planner are its ability to navigate in challenging off-road environments with various terrain types and its fast learning capabilities. By utilizing minimal human demonstration data (5-10 mins), it quickly learns to navigate in a wide array of off-road conditions. The local planner significantly reduces the real world data required to learn human driving preferences. This allows the planner to apply learned behaviors to real-world scenarios without the need for manual fine-tuning, demonstrating quick adjustment and adaptability in off-road autonomous driving technology.

Figures (7)

• Figure 1: Architecture of the classifier. Layers and data are represented by blue and green rectangles respectively. Here $m=21$, Since we have 21 different trajectories in the preference set $S$ (Section \ref{['pref']})
• Figure 2: Top down view of the AirSim environment
• Figure 3: Shows the testing environment. Different scenarios are circled and are numbered from 1-7
• Figure 4: Shows the warthog trajectory (pink curve) in the testing scenarios. All the scenarios are shown by the white ellipses.
• Figure 5: Shows training vs validation loss.