ESA GNC Conference Papers Repository

A new experimental facility for testing of vision-based GNC algorithms for planetary landing
P. Lunghi, M. Ciarambino, M. Marcon, M. Lavagna
Presented at:
Salzburg 2017
Full paper:

The next generation of space exploration missions is going to require an increasing level of autonomy. Telecommunication delays due to interplanetary distances do not allow the ground controllers to undertake prompt reactions to unexpected events. In this framework, autonomous planetary landing is one of the most challenging problems: the duration of the maneuver is relatively short, and in some cases there is no possibility to characterize the landing site from the orbit (i.e. in case of thick atmosphere) with enough detail to ensure the required level of safety. Moreover, most scientifically interesting targets are often placed in hazardous areas. To successfully accomplish pinpoint landing with hazard detection and avoidance, major improvements in Guidance, Navigation and Control are needed: vision-based systems are one of the most promising technologies to achieve the required level of precision and accuracy. Nevertheless, such algorithms need to be carefully tested and validated, in order to ensure the necessary level of robustness. Affordable and repeatable datasets from real mission are scarcely available, often lacking the additional metadata necessary for the computation of a ground-truth solution to compare with the obtained results. Artificially generated images can be a good substitute, but the required level or realism implies the use of computationally intensive, high fidelity rendering algorithms that make difficult closed-loop testing (and impossible real-time hardware-in-the-loop). The use of analog facilities, capable to simulate a landing in a scaled environment, can supply repeatable and controllable data. <p/>This paper presents the design and testing activities of a new experimental facility currently under development at Politecnico di Milano, at the premises of the Aerospace Science and Technology Department (DAER). The system, focused to the simulation of planetary landing, is made up of 4 core subsystems: A <i>terrain diorama</i>, that simulates the planet surface from a visual point of view; A 7 DoF <i>robotic arm</i>, that carries a suite of navigation sensors (mainly a camera, together with possible additional sensors that can be required by the specific GNC algorithm under test); An <i>illumination system</i>, with the role to exclude external light sources and supply a controllable realistic illumination of the scene; A <i>simulation and control</i> <i>unit</i>, that interfaces the algorithms under test with the sensors and the robotic arm. The robotic arm moves the sensor over the diorama, providing stream of images and other measures representative of landing maneuvers; the same diorama can be used at different scale factors, simulating different subphases of a landing. The facility is designed to perform different levels of testing activities: the robotic arm can be moved along predefined trajectories, operating as a simple dataset generator; software-in-the-loop simulations can be performed coupling the GNC algorithms with the simulation of the spacecraft dynamics. Furthermore, the system is compatible with a real-time implementation of the dynamic simulation, in order to allow also hardware-in-the-loop tests. For navigation testing purposes, the knowledge of the position of the camera with respect to the terrain surface is required, with precision and accuracy at least one order of magnitude better than what is expected by the algorithms under test. In practice, the actual value depends also on the maximum scale factor considered. A sub-millimeter precision has been obtained by optical calibration of the diorama, exploiting dense matching techniques. Two different GNC algorithms, developed at DAER, have already been tested: a vision-based navigation system, based on features extraction and tracking, and a hazard detection and target selection system, based on artificial neural networks. The Moon has been taken as representative scenario; preliminary results are shown and discussed.