Development of on-board algorithms to support the navigation of high-speed planetary rovers

The aim of the Master Thesis is to develop, study and validate a path planning algorithm that can support the navigation of rovers for planetary exploration. In particular, the objective is to develop a global path planning algorithm that could be integrated in the rover auto-nav system in order to improve the autonomy and velocity of the planetary rover. Global path planning using grid-based model of the environment is a well-known problem in AI, planning and robotics with a variety of methods and algorithms proposed so far. This work presents a deep-learning approach to the path planning problem. In particular, grid maps, containing information about the traversability of the terrain, are suitable input to modern neural networks, such as convolutional neural networks. The thesis proposes a modern approach, based on the recent advances in deep learning, in order to treat the path planning problem as an Image-to-Image translation problem. Specifically, this thesis applies a Generative Adversarial Network (GAN) and reports the main, obtained results. The GAN network is trained using a wide enough dataset, which contains a set of grid maps, start and goal nodes, and feasible set of paths generated by any well-known algorithm. This kind of approach allows to generalize the path planning problem. The main advantage is that we can use any algorithm, from the simpler to the more complex one, for the generation of training dataset and in such a way the deep learning architecture, trained properly, replaces the necessity to execute the path-planning on-board classic algorithm using a mathematical model that works only with tensor operations. In our tests, in order to generate the paths for the training dataset the well-known and effective algorithm A* is used. The network used, for our application, is a modified version of the Pix2Pix model, introduced by Philip Isola et al. in the paper “Image-to-Image Translation with Conditional Adversarial Networks” (2018) which is a revisitation of the common Conditional Adversarial Network. Using this kind of architecture allows us to influence the generation of the image in such a way that is a plausible translation of the input images, which means having a correct path generated given the generator inputs: environment map and start and goal nodes. We conduct a number of experiments in order to validate the effectiveness of the proposed method, and the results demonstrate how the model adapts well to changes in the position of obstacles in the environment. This work is a part of a research project called SINAV, sponsored by Italian Space Agency. The objective of SINAV is the development of High-Speed Autonomous Navigation Systems for future missions of robotic planetary exploration.