Análise de Terrenos

Three-Dimensional Mapping with Augmented Navigation Cost through Deep Learning

Autonomous navigation for ground robots in unstructured outdoor environments has been the focus of numerous research efforts in Field Robotics in the past few years. External natural environments are especially challenging for ground robots as they exhibit heterogeneous roughness surfaces, including vegetation, pebbles, sand, mud, snow, ice patches, and water puddles. Such surfaces may also be irregular and present different slopes, which significantly increase the difficulty for ground robots to perform their tasks successfully.

A promising route towards solving this challenge is to augment the robot’s maps with rich, informative data streams which would allow the robot to more accurately estimate the navigation cost (difficulty) of traversing different areas. 

Typical approaches used to obtain Navigation Cost Maps such as map an outdoor environment with estimates of the navigation cost for various robot poses. However, most of the works in state-of-the-art literature focus on the response of a single sensor, which potentially constrains the ability to generate accurate navigation costs.

In this paper, we propose a learning-based multi-sensor approach for generating an outdoor terrain cost map, based on the fusion of inertial measurements and LiDAR data. Our approach helps to improve the autonomous navigation of ground robots in an external unstructured environment, as illustrated in Fig. 1.

By combining information provided by different sensory modalities, we can assign navigation costs across a global map. The generated cost maps are combined with traditional path planning approaches to generate optimal paths across the map, optimizing aspects such as traveled distance, time, or energy expenditure. The costs of navigating through an outdoor environment are reduced as in [6]. 

The main contribution is the multimodal representation of unknown terrain. The aforementioned representation is based on the prediction of inertial  measurements from LiDAR data, regarding speed-invariant inertial signals. Our methodology trains a Convolutional Neural Network (CNN) on recorded LiDAR and IMU data, and learns to predict a navigation cost for previously-unseen terrain patches.

From the LiDAR and inertial data, initially a three-dimensional map and the robot localization are computed. From the raw inertial measurement, a speed-invariant inertial data is generated; the roughness level classification is performed and a navigation cost is estimated. Finally, a CNN model is trained from LiDAR and inertial estimations. Based on the trained CNN model, the robot is able to compute the navigation cost for a full map without the necessity or directly observing or driving over it, in contrast to [7], where only inertial sensors are used to estimate navigation costs for unvisited outdoor regions. Furthermore, the resulting map provides an accurate representation of regions and can be directly used as a valuable input into path planning algorithms which can increase safety and reduce energy expenditure.

Experiments on real-world scenarios show the efficacy of the inertial speed-invariant transformation, minimizing the effect of variable speed on inertial data  acquisition. The experiments also show that the proposed roughness level classification process achieved a mean accuracy of over 84%, even considering different environments, speeds, and terrain features. Additionally, the experiments suggest that the obtained maps are accurate and applicable in the navigation process.

Publications:

OLIVEIRA, FELIPE G.; NETO, ARMANDO A. ; HOWARD, DAVID ; BORGES, PAULO ; CAMPOS, MARIO F. M. ; MACHARET, DOUGLAS G. . Three-Dimensional Mapping with Augmented Navigation Cost through Deep Learning. JOURNAL OF INTELLIGENT & ROBOTIC SYSTEMS, v. 101, p. 50, 2021.

OLIVEIRA, JUAN M. A. ; MACHARET, DOUGLAS G. ; OLIVEIRA, FELIPE G. . Roughness Level Classification using Inertial Data for Wheeled Robots in Outdoor Terrains. In: 2023 Latin American Robotics Symposium (LARS), 2023 Brazilian Symposium on Robotics (SBR), and 2023 Workshop on Robotics in Education (WRE), 2023, Salvador. 2023 Latin American Robotics Symposium (LARS), 2023 Brazilian Symposium on Robotics (SBR), and 2023 Workshop on Robotics in Education (WRE), 2023. p. 343.

OLIVEIRA, FELIPE G.; NETO, ARMANDO ALVES ; BORGES, PAULO ; CAMPOS, MARIO F. M. ; MACHARET, DOUGLAS G. . Augmented Vector Field Navigation Cost Mapping using Inertial Sensors. In: 2019 19th International Conference on Advanced Robotics (ICAR), 2019, Belo Horizonte. 2019 19th International Conference on Advanced Robotics (ICAR), 2019. p. 388.

OLIVEIRA, FELIPE G.; SANTOS, ELERSON R. S. ; NETO, ARMANDO ALVES ; CAMPOS, MARIO F. M. ; MACHARET, DOUGLAS G. . Speed-invariant terrain roughness classification and control based on inertial sensors. In: 2017 Latin American Robotics Symposium (LARS) and 2017 Brazilian Symposium on Robotics (SBR), 2017, Curitiba. 2017 Latin American Robotics Symposium (LARS) and 2017 Brazilian Symposium on Robotics (SBR), 2017. p. 1.

Acknowledgments:

This work was developed with the support of Conselho Nacional de Desenvolvimento Cientı́fico e Tecnológico (CNPq), Coordenação de Aperfeiçoamento de Pessoal de Nı́vel Superior (CAPES), Fundação de Amparo à Pesquisa do Estado de Minas Gerais (FAPEMIG), Fundação de Amparo à Pesquisa do Estado do Amazonas (FAPEAM) and Commonwealth Scientific and Industrial Research Organisation (CSIRO).

Team:

Felipe G. Oliveira, Adjunct Professor at Universidade Federal do Amazonas (UFAM)

Armando A. Neto, Associate Professor at Universidade Federal de Minas Gerais (UFMG)

David Howard, Researcher at Commonwealth Scientific and Industrial Research Organisation (CSIRO)

Paulo Borges, Researcher at Commonwealth Scientific and Industrial Research Organisation (CSIRO)

Mario F. M. Campos, Titular Professor at Universidade Federal de Minas Gerais (UFMG)

Douglas G. Macharet, Associate Professor at Universidade Federal de Minas Gerais (UFMG)