Abstract
Mobile robots can be applied in a variety of tasks requiring them to traverse rough, unstructured terrains, such as planetary exploration, mining, search and rescue activities, nuclear plant inspections, and agriculture. The large scale of these environments necessitates careful consideration and management of resources in the planning of paths. In this context, the driving energy consumption plays a crucial role in traversing large distances and in the successful achievement of objectives without the risk of stranding the platform out in the field. However, complex terrain interactions with unstructured geometries, unknown soil characteristics, and computing time constraints often limit the accuracy of the driving energy estimations. These challenges need to be addressed before autonomous mobile robots are ready for deployment in challenging, off-road terrains.
This thesis seeks to tackle some of these key issues that impact the autonomy of robotic platforms in off-road environments. Particularly, this work focuses on (1) how the driving energy consumption can be accurately predicted and (2) how energy-efficient paths can be rapidly computed for mobile robots traversing off-road terrains with unstructured geometries and unknown soil characteristics. The key contribution of this thesis is an energy-aware guidance framework that leverages recent advancements in computing and deep learning solutions to address the two above-mentioned problems. Among the novelties of the proposed framework, it is remarked (1) leveraging convolutional 1D neural networks to analyse the complex wheel-terrain interaction over highly unstructured terrain geometries (Chapter 3), (2) adopting a meta-learning approach, based on the online observation of small numbers of proprioceptive and exteroceptive measurements, to capture the effect on the driving energy of unknown soil properties (Chapter 4), (3) combining convolutional 1D neural networks and meta-learning to provide driving energy estimates over terrains with both unstructured geometries and unknown soil properties (Chapter 5), (4) re-framing the deep neural network architecture in a probabilistic fashion to provide uncertainty-aware considerations to enhance the robot safety (Chapter 6), and (5) proposing different solutions, in each chapter, to integrate the neural-network-based energy predictors into a graph-search path planning algorithm so as to enable energy-aware path planning.
The proposed framework is tested using two state-of-the-art 3D-body dynamic simulators running on standard desktop computers. The former enables a realistic modelling of a robotic platform and its dynamic interaction with environments characterised by complex, unstructured geometries. The latter, in addition to unstructured geometries, can model several types of terrains with realistic rigid and deformable soil properties. The numerical values obtained by running the simulators are used to train the proposed deep learning models and, upon training completion, to compare their prediction performance against alternative driving energy prediction solutions. Furthermore, the prediction models are integrated, at each step, into a graph search path planning algorithm and their planning performance are analysed in terms of path optimality and computational efficiency.
The numerical results show evidence of the benefits of the proposed framework when traversing challenging, off-road terrains characterised by highly unstructured geometries and unknown soil properties. In terms of driving energy prediction, a considerable improvement of the performance is obtained over alternative state-of-the-art methods. Moreover, the probabilistic reformulation further improves the reliability of the system as it can yield more informative uncertainty-aware estimations that can be crucial to the safety of the robot when traversing challenging terrains. Regarding the planning performance, the proposed framework can compute more energy-efficient paths than alternative state-of-the-art solutions while preserving similar computational efficiency. Furthermore, it sensibly reduces the computational time compared to highly accurate, but computationally expensive and difficult-to-tune, path planning approaches based on physics-based energy models. Therefore, the proposed framework can be of interest to many applications of autonomous mobile robots in challenging off-road terrains, for which providing accurate driving energy estimations and energy-efficient paths are crucial tasks.