Safe Long-Range Travel for Planetary Rovers through Forward Sensing

Transcription

1 Safe Long-Range Travel for Planetary Rovers through Forward Sensing 12 th ESA Workshop on Advanced Space Technologies for Robotics & Automation Yashodhan Nevatia Space Applications Services May 16 th, 2013

2 This research project Forward Acquisition of Soil and Terrain data for Exploration Rover (FASTER), running from 2011 to 2014, is supported by the European Commission through the SPACE theme of the FP7 Programme under Grant Agreement Yashodhan Nevatia Jeremi Gancet Francois Bulens Roland Sonsalla Martin Frische Thomas Voegele Thilo Kaupisch* Yang Gao Said Al-Milli Rajeev Kandiyil Chakravarthini Saaj Marcus Matthews Brian Yeomans Stephen Ransom Lutz Richter Elie Allouis Krzysztof Skocki *Thilo Kaupisch was affiliated with DFKI at the time part of the work reported here was performed, and is currently affiliated with DLR ev. 2

3 Motivation Advances in autonomy for terrestrial robots DARPA Grand Challenges (2004, 2005, 2007) Google Autonomous Car Etc Planetary rovers operations Minimal risk Largely planned by operators so far Analysis, simulation based on telemetry Short turn around time Limited to line of sight operations Still dangerous MER Spirit Future missions require greater rover capabilities! ExoMars mission length of ~ 9 months Mars Sample Fetch Rover - ~ 15 km traversal Photos: NASA/JPL-Caltech 3

4 Objectives Autonomous operations allowing safe long range traversal Forward sensing of terrain trafficability Allow detection of hazards with minimal risk to mission rover Components as mission add on System components Scout rover Soil sensors Primary rover sensors Scout rover sensors Remote sensing Cooperative autonomy 4

5 Operation Concept Primarily to pure traversal operations Minimal or no science during traverse Ground Planning Phase Identification and specification of traverse command On Board Command Procedures Global Path Planning Waypoint Traversal 5

6 Ground Planning Phase Preparation for a single traverse command Identification of target location Can be more than one sol away > 200 m Identification of potential paths Using orbiter data Path as set of waypoints Straight line between waypoints Estimated (directional) cost of traversal Multiple, interconnected paths Paths encoded as a directional graph Waypoints as nodes, costs as edge weights 6

7 Global Path Planning Start Traverse G Target Location D E B C A O Current Location (Start) 7

8 Global Path Planning Start Traverse G D E Find best path Simple graph search Best path: (O, A, C, E, G) B A C O 8

9 Global Path Planning Start Traverse G D E Find best path Waypoint Traversal Path found Y B A C Sub procedure Traversal to next waypoint O 9

10 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost B A C N Target reached O 10

11 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost B A C N Target reached O 11

12 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached O 12

13 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached O 13

14 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached O 14

15 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached O 15

16 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached O 16

17 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached O 17

18 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found Y Success Y Update edge cost N B Remove edge A C N Target reached Y Traverse success O 18

19 Global Path Planning Start Traverse G Find best path Waypoint Traversal D E Path found N Y Success Y Update edge cost N B Remove edge A C Traverse fail N Target reached Y Traverse success O 19

20 Waypoint Traversal Similar to behaviours proposed by Volpe et. al. [1] Motion-to-Goal and Boundary-Following Waypoint Rovers turn towards the next waypoint Build elevation map 20

21 Waypoint Traversal Similar to behaviours proposed by Volpe et. al. [1] Motion-to-Goal and Boundary-Following Waypoint Rovers turn towards the next waypoint Build elevation map Remote soil sensing input for rocks 21

22 Waypoint Traversal Similar to behaviours proposed by Volpe et. al. [1] Motion-to-Goal and Boundary-Following Waypoint Rovers turn towards the next waypoint Build elevation map Remote soil sensing input for rocks Artificial target if needed 22

23 Waypoint Traversal Similar to behaviours proposed by Volpe et. al. [1] Motion-to-Goal and Boundary-Following Waypoint Rovers turn towards the next waypoint Build elevation map Remote soil sensing input for rocks Artificial target if needed Path planning 23

24 Waypoint Traversal Similar to behaviours proposed by Volpe et. al. [1] Motion-to-Goal and Boundary-Following Waypoint Rovers turn towards the next waypoint Build elevation map Remote soil sensing input for rocks Artificial target if needed Path planning No path Obstacle circumnavigation Rovers allowed to rotate slightly 24

25 Waypoint Traversal Scout begins forward sensing Path updates at scout location as needed (eg. hazard detection) Extension of elevation map on reaching end of path Constrained to within line of sight of primary rover Primary rover traversal Path update from primary rover location on detection of hazard Scout returns to primary rover location 25

26 Waypoint Traversal Failure No path found, rovers at waypoint No new nodes needed B A 26

27 Waypoint Traversal Failure No path found, rovers at waypoint No new nodes needed Hazard detection by scout Addition of two connected nodes Scout to primary, or vice versa A2 B A1 A 27

28 Waypoint Traversal Failure No path found, rovers at waypoint No new nodes needed Hazard detection by scout Addition of two connected nodes Scout to primary, or vice versa Hazard detection by primary rover Single new node Scout to primary rover B A1 A 28

29 Waypoint Traversal Failure No path found, rovers at waypoint No new nodes needed Hazard detection by scout Addition of two connected nodes Scout to primary, or vice versa Hazard detection by primary rover Single new node Scout to primary rover New nodes are connected Last waypoint Other nodes in line of sight? C B A1 A 29

30 Cooperative Autonomy & Software E3 level of autonomy Adaptive mission operations on-board Partial support for E4 level of autonomy Operation concept provides for goal based traversal Scout Rover Minimal autonomy Primary rover Path following Return to primary Proprioception Subsystem architecture based on G en om with ROS modules 30

31 Primary Rover Software Architecture 31

32 Representative Subsystems Task Execution Controllers Simple on board procedure engines Predefined procedures Health Management Offline analysis based fault detection and recovery Data Management Simplified data repository supporting pull/push Communication Wireless communication with scout Simulated communication with ground 32

33 Representative Subsystems Locomotion Controller Implements drive commands TCP interface for Bridget Controller Sensor Managers Point Grey XB3 Stereo Camera XSens MTi IMU 33

34 SSS Software Chain Interface with soil sensing system Traversability classification Based on outputs of multiple sensors Rock tracking via Remote Sensing 34

35 Task Planner Planning and monitoring high level tasks Symbolic planner Hierarchical Task Networks Domain configurable, multi agent framework [2] Generation of interleaved plans Multiple, concurrent action sequences Query functional modules for cost estimates Re-planning, including contingency actions Validation of plans with regard to resources 35

36 Scout Localization Tracking 6DOF scout pose in images Combined with scout proprioception SURF feature based: 3D descriptor database Generation of image-model correspondences Marker based: ARToolkit for benchmarking Custom marker configurations Single Marker Multi Marker 36

37 GNC: Localization Wheel and/or Inertial Odometry Visual Odometry Rock Tracking Output from Remote sensing Sparse feature SLAM being investigated Map matching Detailed local elevation map built by rover Low resolution orbiter height map Bounds for odometry error 37

38 GNC: Mapping & Path Planning Elevation maps based on stereo camera images Filtering Iterative Closest Point for 3D data fusion D* based local path planning, trajectory fitting Remote Soil Sensing Scout Localization Primary Rover Stereo Images Point Cloud Point Cloud Merging (ICP) Scout Rover Stereo Images Point Cloud DEM Generation Path Planning 38

39 System Validation Simulation Autonomy Field Trials Rapid development & testing Gazebo Multi-robot outdoor simulation Open Dynamics Engine DARPA Robotics Challenge HiRISE imagery from MRO Integrated Field Trials 39

40 System Validation Simulation Iterative validation of algorithms Preliminary integration with scout Autonomy Field Trials Realistic datasets Clearpath Husky A200 Integrated Field Trials Beez Quarry, Belgium 40

41 System Validation Simulation Autonomy Field Trials Integrated system validation Soil sensors Bridget locomotion breadboard Field trials at Stevenage Second half of 2014 Integrated Field Trials Photo: PRoVisG Field Trial,

42 Thank you! Yashodhan Nevatia Work Package Leader Space Applications Services Chakravarthini Saaj Technical Manager University of Surrey Thomas Vögele Project Coordinator German Research Center for Artificial Intelligence

43 References [1] Volpe, R., Estlin, T., Laubach, S., Olson, C., & Balaram, J. (2000). Enhanced mars rover navigation techniques. In Proc. IEEE International Conference on Robotics and Automation, 2000 (ICRA'00), IEEE, Vol. 1, pp [2] R. Kandiyil and Y. Gao, A Generic Doman Configurable Planner using HTN for Autonomous Multi-Agent Space System, in Proc. 11th International Symposium on Artificial Intelligence, Robotics and Automation in Space, Turin, Italy,

44 MSR Mission Outline Mission Surface Phases Activities Duration Units Post Landing Checkout Comms establishment & HK status 4 sols Preparation for Departure from Landing Point Sample Cache Acquisition Cache transfer to MSR Lander Contingencies Initial Post landing checkout 4 sols Egress 4 sols Post-landing checkout duration 12 sols SFR commissioning 5 sols Landing Site local exploration 2 sols Pre-excursion duration (sols) 7 sols Identification of target location 1 sols travel to target location 3 sols Verification & confirmation of target 2 sols Sample Acquisition 1 sols Sample Cache Acquisition Duration 7 sols Identification of target location 1 sols travel to target location 2 sols cache Transfer 1 sols Cache Transfer to MSR Duration 4 sols Conjunction Allocation (no ops) 20 sols Dust storm or operational contingency 20 sols Total Contingency 40 sols Total Mission Operations duration 70 sols Nominal Mission Duration 180 sols Duration left for traverse 110 sols Traverse Distance 15 km Traverse Margin 1.3 Minimum Rover speed 177 m/sol 44

45 Scout Forward Sensing Three scenarios considered with different parameters

46 Scout Forward Sensing turns Measurement time of 60 s Step pattern is too slow to achieve desired daily traverse Scenario 3 chosen as baseline Most plausible in terms of mission size and complexity 46

47 Extended Operations Exploration or scientific tasks Very useful, but considered out of scope Long Term Scouting beyond range of Primary Rover sensors/communication delays in case of wrong trafficability from scout sensors Increased requirements for Scout Fully autonomous operation Self localization and task planning Increased power requirements Increased mission complexity Relative localization between scout path and primary rover Recharging of scout batteries 47

48 Remote Sensing Semantic feature detection rather than pixel-based feature detection Three approaches are currently being investigated: o Supervised machine learning classifiers (SVM & Boosted LDA): Stage-1 (Learning) Stage-2 Classification Positive Patch F(x) = 1 Negative Patch F(x) =

49 Remote Sensing Blob Detection, classification, and tracking Saliency Detection, classification, and tracking Stage-1 Stage-2 49