Reconstructing occluded Elevation Information in Terrain Maps with Self-supervised Learning (2109.07150v2)

Abstract: Accurate and complete terrain maps enhance the awareness of autonomous robots and enable safe and optimal path planning. Rocks and topography often create occlusions and lead to missing elevation information in the Digital Elevation Map (DEM). Currently, these occluded areas are either fully avoided during motion planning or the missing values in the elevation map are filled-in using traditional interpolation, diffusion or patch-matching techniques. These methods cannot leverage the high-level terrain characteristics and the geometric constraints of line of sight we humans use intuitively to predict occluded areas. We introduce a self-supervised learning approach capable of training on real-world data without a need for ground-truth information to reconstruct the occluded areas in the DEMs. We accomplish this by adding artificial occlusion to the incomplete elevation maps constructed on a real robot by performing ray casting. We first evaluate a supervised learning approach on synthetic data for which we have the full ground-truth available and subsequently move to several real-world datasets. These real-world datasets were recorded during exploration of both structured and unstructured terrain with a legged robot, and additionally in a planetary scenario on Lunar analogue terrain. We state a significant improvement compared to the baseline methods both on synthetic terrain and for the real-world datasets. Our neural network is able to run in real-time on both CPU and GPU with suitable sampling rates for autonomous ground robots. We motivate the applicability of reconstructing occlusion in elevation maps with preliminary motion planning experiments.

• The paper introduces a novel self-supervised framework that uses a U-Net architecture with ray casting to overcome occlusions in digital elevation maps.
• It achieves significant error reductions ranging from 52% to 82% across both synthetic and real-world datasets.
• The enhanced reconstruction of elevation data improves autonomous navigation by enabling more effective traversability analysis in complex terrains.

Reconstructing Occluded Elevation Information in Terrain Maps with Self-supervised Learning

The paper "Reconstructing Occluded Elevation Information in Terrain Maps with Self-supervised Learning" addresses a significant challenge in robotic navigation—dealing with incomplete digital elevation maps (DEMs) resulting from occlusions and sensor limitations. The research introduces a novel self-supervised learning methodology to accurately reconstruct occluded areas in these maps, enhancing autonomous robots' operational capabilities in various terrains, including both structured and unstructured environments.

Methodology and Approach

The core contribution of this paper lies in its self-supervised learning framework that circumvents the need for complete ground-truth data, which is often unavailable in real-world settings. The approach employs a U-Net-like neural network architecture for processing the elevation data. Unlike conventional supervised methods, this approach leverages artificial occlusion via ray casting—a technique that mimics natural obstruction from a randomly selected vantage point. This process creates a training dataset where the added occlusion can serve as a controlled experiment to train the neural net effectively.

The model is meticulously designed, incorporating both Mean Squared Error (MSE) and Total Variation (TV) losses to ensure smooth and realistic reconstructions. Notably, the proposed network differentiates between occluded and non-occluded areas by emphasizing reconstruction loss in the occluded regions, thereby training more effectively on realistic terrain data.

Empirical Evaluation

The paper provides a comprehensive evaluation of the proposed method across multiple synthetic and real-world datasets, including the gonzen mine with the ANYmal legged robot and a dataset from a planetary scenario involving the Heavy-Duty Planetary Rover. The results indicate a substantial improvement over traditional baseline methods, with reductions in the MSE error ranging from 52% to 82% across different datasets. This substantial numerical gain highlights the algorithm's robustness and adaptability to various terrain complexities.

Implications and Future Directions

Practically, this research offers a significant upgrade to the robustness and reliability of autonomous navigation systems by enhancing their capability to operate on incomplete data. Robots equipped with this technology can exhibit improved traversability analysis, paving the way for more efficient path planning even in previously intractable environments.

From a theoretical perspective, this work pioneers a new trajectory in self-supervised learning applications for robotic perception, particularly emphasizing the unorthodox use of ray casting for training data generation. It sets a precedent for further exploration into active learning systems capable of dynamically enhancing their datasets.

For future developments, the paper suggests the integration of generative adversarial networks (GANs) to generate even more realistic occlusion patterns. Additionally, incorporating perceptual losses, typically reliant on complete datasets, could further refine the quality of the reconstructed DEMs. Moreover, estimating model and data uncertainty would be an invaluable enhancement, aiding the deployment of these models in safety-critical applications like Mars rovers or subterranean exploration missions.

In conclusion, this work represents an advancement in robotic terrain navigation and sets a framework for future research focused on overcoming data limitations through novel machine learning strategies.