Proprioception Is All You Need: Terrain Classification for Boreal Forests (2403.16877v2)

Abstract: Recent works in field robotics highlighted the importance of resiliency against different types of terrains. Boreal forests, in particular, are home to many mobility-impeding terrains that should be considered for off-road autonomous navigation. Also, being one of the largest land biomes on Earth, boreal forests are an area where autonomous vehicles are expected to become increasingly common. In this paper, we address this issue by introducing BorealTC, a publicly available dataset for proprioceptive-based terrain classification (TC). Recorded with a Husky A200, our dataset contains 116 min of Inertial Measurement Unit (IMU), motor current, and wheel odometry data, focusing on typical boreal forest terrains, notably snow, ice, and silty loam. Combining our dataset with another dataset from the state-of-the-art, we evaluate both a Convolutional Neural Network (CNN) and the novel state space model (SSM)-based Mamba architecture on a TC task. Interestingly, we show that while CNN outperforms Mamba on each separate dataset, Mamba achieves greater accuracy when trained on a combination of both. In addition, we demonstrate that Mamba's learning capacity is greater than a CNN for increasing amounts of data. We show that the combination of two TC datasets yields a latent space that can be interpreted with the properties of the terrains. We also discuss the implications of merging datasets on classification. Our source code and dataset are publicly available online: https://github.com/norlab-ulaval/BorealTC.

• The paper introduces the BorealTC dataset from proprioceptive sensors to enable robust terrain classification in boreal forests.
• The paper compares a CNN and a state-space-based Mamba model, finding that Mamba scales better and delivers higher accuracy on combined datasets.
• The study highlights proprioception’s resilience over traditional sensors in harsh boreal conditions, paving the way for improved autonomous navigation.

Proprioception Is All You Need: Terrain Classification for Boreal Forests

The paper "Proprioception Is All You Need: Terrain Classification for Boreal Forests" offers a comprehensive exploration of terrain classification (TC) using proprioceptive sensors for autonomous navigation in boreal forests. This biome, characterized by challenging terrain conditions such as snow, ice, and silty loam, demands robust navigation capabilities from autonomous vehicles due to its vastness and the increasing prevalence of such vehicles in these environments. The authors present a novel dataset, BorealTC, showcasing terrain data captured via proprioceptive sensors on a Husky A200 robotic platform to address this need.

Methodological Overview

The paper employs two contrasting baseline models for terrain classification: a Convolutional Neural Network (CNN) and the Mamba architecture, which is based on state space models (SSM). The CNN model processes spectrograms of proprioceptive sensor data and has been optimized with techniques such as windowing functions to enhance its performance. In contrast, the Mamba model leverages its state space architecture to manage raw sequential data effectively, providing a scalable solution for lengthy sequences.

The research emphasizes the relative advantages of proprioceptive sensors—such as IMUs, odometry, and motor currents—which are less susceptible to environmental constraints like lighting and occlusion, issues that typically impair exteroceptive sensors like cameras and LIDARs in the complex visual conditions of boreal forests.

Numerical Results and Analysis

Through comparative analysis, the paper highlights that while the CNN outperformed Mamba on the individual Vulpi and BorealTC datasets, Mamba demonstrated superior accuracy when both datasets were combined. This outcome indicates Mamba’s potential efficacy in scenarios involving vast, complex training sets, aligning with the hypothesis that performance is significantly data-limited.

Quantitative assessment reveals robust precision and recall scores across terrains, with notable accuracies on challenging surfaces, particularly for larger datasets where the Mamba architecture capitalizes on its scalability advantages. This highlights a meaningful finding: the efficacy of the Mamba model improves with increased data, which suggests a potential shift in favor of SSM-based models in data-intensive scenarios.

Implications and Future Directions

The implications of this work are twofold. Practically, it underscores the viability of proprioceptive-based TC systems for reliable autonomous navigation across seasons in boreal forests. Theoretically, it suggests a trajectory for future models to explore further state space-based architectures to handle multi-seasonal, complex data environments effectively. Merging different datasets without significant performance detriments indicates pathways for future knowledge transfer between various terrain datasets, which is crucial for scaling autonomous systems to navigate diverse and dynamic environments.

For evolving AI and robotics, this research opens up avenues for refining proprioceptive sensor utilization and architecting more adaptable models capable of managing domain shifts and varied data input efficiently. Future work may focus on the standardization of TC data acquisition and further exploration of domain adaptation techniques to improve model robustness across diverse settings. Additionally, investigating other state space architectures or hybrid models, which combine CNNs with SSMs, could offer substantial improvements in TC, especially for environments as variable and demanding as boreal forests.

Overall, the paper crucially enriches the discourse on terrain classification methodologies, offering not only novel insights but also a valuable dataset for the broader TC research community.