Acta Futura Issue 9 AI in Space Workshop at IJCAI 2013 Advanced Concepts Team http://www.esa.int/act Publication: Acta Futura, Issue 9 (2014) Editor-in-chief: Duncan James Barker Associate editors: Dario Izzo Jose M. Llorens Montolio Pacôme Delva Francesco Biscani Camilla Pandolfi Guido de Croon Luís F. Simões Daniel Hennes Editorial assistant: Zivile Dalikaite Published and distributed by: Advanced Concepts Team ESTEC, European Space Research and Technology Centre 2201 AZ, Noordwijk e Netherlands www.esa.int/actafutura Fax: +31 71 565 8018 Cover image: Four-wheel skid-steered robot used in data collection experiments from Kohler et al. in this issue. ISSN: 2309-1940 D.O.I.:10.2420/ACT-BOK-AF Copyright ©2014- ESA Advanced Concepts Team Typeset with X TE EX Contents Foreword 7 Sensor Fault Detection and Compensation in Lunar/Planetary Robot Missions Using Time-Series Prediction Based on Machine Learning 9 Tim Köhler, Elmar Berghöfer, Christian Rauch and Frank Kirchner Planning Mars Rovers with Hierarchical Timeline Networks 21 Juan M. Delfa Victoria, Simone Fratini, Nicola Policella, Oskar von Stryk, Yang Gao and Alessandro Donati Onboard Autonomous Response for UAVSAR as a Demonstration Platform for Future Space-Based Radar Missions 31 Joshua Doubleday, Steve Chien, Yunling Lou, Duane Clark and Ron Muellerschoen Lunar Crater Identification from Machine Learning Perspective 41 Chak Pong Chung, Cheuk On Chung, and Chit Hong Yam GPU Accelerated Genetic Algorithm for Low-thrust GEO Transfer Maneuvers 49 Kai Yu and Ming Xu How Ants Can Manage Your Satellites 59 Claudio Iacopino, Phil Palmer, Nicola Policella, Alessandro Donati and Andy Brewer Pattern-Based Modeling for Timeline Planning in Space Domains 73 Simone Fratini, Nicola Policella and Alessandro Donati Procedural Onboard Science Autonomy for Primitive Bodies Exploration 83 Steve Chien, Gregg Rabideau, Dero Gharibian, David ompson, Kiri Wagstaff, Brian Bue and Julie Castillo-Rogez Space Hopper: a Serious Game Crowdsourcing the Design of Interplanetary Trajectories 93 Wiktor Piotrowski, Marcus Märtens, Dario Izzo and Daniel Hennes 5 6 Foreword I is making significant inroads in the space sector. Artifi- cial Intelligence (AI) systems are contributing to numerous space missions includ- A ing the Hubble Space Telescope, Mars Exploration Rovers, the International Space Station, and Mars Express. Many different areas of AI are already, or may be in the near future, of particular interest from a space applications point of view. ese include topics as diverse as: intelligent search and optimization methods in aerospace applications, im- age analysis for guidance navigation and control, autonomous exploration of interplanetary and planetary environments, intelligent algorithms for fault identification, as well as data mining and visual presentation of large data sets. is issue contains selected papers from the 2013 AI in Space workshop, held as a satel- lite event at the 23 International Conference on Artificial Intelligence (IJCAI) in Beijing, China. Now in its fourth edition, this workshop was co-organized by the Advanced Con- cepts Team of the European Space Agency and the Artificial Intelligence Group of NASA’s Jet Propulsion Laboratory. e goal of the AI in Space workshop is to highlight the most recent AI applications related to space, encourage collaboration between the two areas, and to provide an overview of current research. e following 9 papers were presented at the AI in Space workshop and selected for revised submission. As a result, this issue provides an interesting compilation of current research directions in the area of artificial intelligence applied to the space domain. Sensor Fault Detection and Compensation in Lunar/Planetary Robot Missions Using Time-Series Prediction Based on Machine Learning by Tim Köhler et al. introduces a system that detects sensor drop outs and compensates missing sensor signals. Planning Mars Rovers with Hi- erarchical Timeline Networks by Juan M. Delfa Victoria et al. presents a heuristic planner to solve highly constrained temporal problems for autonomous on-ground and on-board planning. Onboard Autonomous Response for UAVSAR as a Demonstration Platform for Fu- ture Space-Based Radar Missions by Joshua Doubleday et al. discusses a system for onboard processing and interpretation of radar data as well as autonomous retasking of the vehi- cle and instrument. Lunar Crater Identification from Machine Learning Perspective by Chak Pong Chung, Cheuk On Chung, and Chit Hong Yam explores different methods for lunar crater detection based on digital elevation model data provided by the Chinese Chang’E-1 and Chang’E-2 spacecrafts. GPU Accelerated Genetic Algorithm for Low-thrust GEO Trans- fer Maneuvers by Kai Yu and Ming Xu presents a GPU based implementation of a novel genetic algorithm for low-thrust trajectory design. How Ants Can Manage Your Satellites by Claudio Iacopino et al. discusses the design of an innovative ground-based automated planning and scheduling system based on ant colony optimization. Pattern-Based Modeling for Timeline Planning in Space Domains by Simone Fratini, Nicola Policella and Alessan- dro Donati describes a system for pattern-based modeling to bridge the gap between AI planning languages and the current practice in the space operational context. Procedural Onboard Science Autonomy for Primitive Bodies Exploration by Steve Chien et al. presents a close loop autonomous approach for onboard science target detection and response for time constraint missions to primitive bodies. Space Hopper: a Serious Game Crowdsourcing the Design of Interplanetary Trajectories by Wiktor Piotrowski et al. introduces an online crowdsourcing experiment that aims to improve automated trajectory design. Daniel Hennes and Dario Izzo (Associate editors) 7 8 Acta Futura 9 (2014) 9-20 Acta DOI: 10.2420/AF09.2014.9 Futura Sensor Fault Detection and Compensation in Lunar/Planetary Robot Missions Using Time-Series Prediction Based on Machine Learning T K*¹, E B¹, C R² F K¹ ¹ DFKI GmbH, Robotics Innovation Center, Robert-Hooke-Straße 5, 28359 Bremen, Germany ² University of Bremen, Robotics Research Group, Robert-Hooke-Straße 5, 28359 Bremen, Germany Abstract. Mobile robots operating in a lunar 1 Introduction or planetary space mission can usually neither be repaired from nor brought back to earth. In case Robotic applications in space have a high demand on of sensor damages or drop outs an overhaul of the the system’s reliability: there are only short communi- hardware leading to a properly working sensor is cation windows, a high latency in communication, and not possible in most cases. Instead, the system there is nearly no way of recovering a system when it is has to continue working as reliable as possible. In in a fault state. Different hierarchical behavior archi- the special case of an autonomous space robot this tectures were developed (e.g., [2]) to enable the robot means that the robot, first, needs to detect a sen- sor drop out automatically. Second, the missing to carry out high-level plans and to supervise the be- sensor signal needs to be compensated. In typical havior execution. However, faults in the sensor hard- mobile robot setups this is possible by using other ware would still be a critical problem in plan execution sensor modalities. Presented is a method to de- – although there might be several cases of single sen- tect single sensor faults by model-based predictions sor drop-outs that could be compensated. For a typi- covering multiple sensor modalities. e methods cal lunar or planetary space mission, an example could are learned using one of two methods: tested and be a wheeled mobile robot with a three-axes gyroscope, compared are a multi-layer perceptron (MLP) and a camera or laser range finder, and wheel encoders. A a “Neural Gas (NG)” vector quantization method. malfunctioning single sensor in such a setup could be e test use case is a turning skid-steered robot compensated (at least partly) by a combination of the with four different sensor modalities (velocities left motor commands and the data obtained from the other and right wheels, gyroscope Z-axis, and horizontal optical flow). With the collected training and test sensors. data the model predictions turns out to be accurate By learning the correlation between actions executed enough for the purpose of a sensor fault detection. by the robot and the corresponding sensor feedback, a Moreover, by using learned models a compensation prediction of sensor signals could be possible. us, a in case of a sensor fault can be possible. model of the motor-to-sensor relation is learned which is used to generate expected sensor values when execut- *Corresponding author. E-mail: [email protected] ing a learned action again. Such a model is specific for 9 Acta Futura 9 (2014) / 9-20 Köhler, T. et al. the learned action (or actions) and might depend on cer- is examined. tain properties of the environment. A model-based pre- e first method used to train the prediction model diction with three different methods of model genera- is called “Neural Gas”. Neural gas (NG, Martinetz et tion and with an application of detecting environmen- al. 1993, [6]) is a vector quantization method which tal conditions rather than sensor failures is presented by preserves the structure of the action-perception relation Rauch et al. [7]. and does not fit a functional description to this relation. Comparing the prediction error (between the learned, is especially can be useful if the prediction model has expected state and the actual sensor feedback) of several to cover different situations where the sensor feedback sensor modalities could lead to an identification of sin- differs depending on environmental properties only (i.e. gle sensor drop-outs. Furthermore, if a sensor failure if ambiguities exist in one modality). is identified, a compensation could be possible by us- A second method used in comparison to the NG is ing the predicted sensor values instead of the measured the “Multi Layer Perceptron” (MLP) (Rummelhart et ones.