from cars to mars – applying autonomous vehicle ......from cars to mars – applying autonomous...
TRANSCRIPT
From Cars to Mars – applying Autonomous
Vehicle Navigation Methods to a Space Rover
Mission
Dr Brian Yeomans
Oxford Robotics Institute
http://ori.ox.ac.uk/
Presentation
Outline
● About ORI
● The MURFI 2016 mission
● Dub4 and Experienced Based Navigation
● Monocular navigation and localisation
● Dense Reconstructions
● Conclusions and Next Steps
ORI’s mission Develop technology that allows machines to ask and answer:
“Where am I?” “What surrounds me?” “What should I do next?”
ORI Core Tech
Key elements of our autonomy system: • Divides into several different topics • All platforms share a common
codebase • Targets for integrated system:
• Infrastructure-Free • Operate over Large Scales • Low Cost
Unmanned Platforms
Manned Platforms
Sensors ● Cameras (Stereo and
mono)
● Lidar (3D and 2D)
● Radar
● Ultrasonics
● INS
What do we do with our data?
● Processing... ○ Compression, Indexing,
Visual Odometry
○ Scene Priors, FABMAP
Loop Closures, Global
Optimisation, Map Building
○ BORG, YOLO, Global Map
Database, ...
● Sharing... ○ http://www.ori.ox.ac.uk/datasets/
○ RobotCar Dataset
○ FABMAP Eyensham 70k Dataset
○ New College Vision and Laser
Dataset, ...
● Experiments... ○ Testing for navigation,
planning, ...
○ Training for perception &
planning
○ Simulation for system
design and testing
○ ...
The MURFI 2016 Mission
● Run by UKSA in collaboration with CSA in
two locations, Utah, USA and Harwell, UK
● ExoMars like analogue mission designed
to demonstrate mission capability
MURFI Rover Hardware and
Sensors
4 wheel steering, passive
dynamic suspension
Bumblebee XB3 stereo
camera Long range Wifi (1km +)
Mast mounted PanCAM
(Univ of Aberystwyth)
3D LIDAR Monocular camera
Other instruments as required
MURFI Sol drives
● Planned tracks simple - linear drives linked by
point turns
● Egomotion of Rover generated using Oxford VO
● Differential controller keeps Rover on track Image credit MURFI team
Oxford Visual Odometry
● Flying on
ExoMars
Dub4 / EBN
● “Wherever, Whenever, Whatever the Weather” (dub4)
● Teach and Repeat
● Experience stores visual snapshots plus relative coordinates between
snapshots
● More drives -> more experiences. Improve localisation irrespective of
viewpoint, illumination, weather
● Mapping:
○ Uses Oxford VO to estimate trajectory
○ Each stereo image pair creates a bank of landmarks and the 6-
DoF transform from the previous image pair
○ Stored in graph where nodes are the landmarks and the 6-DoF
transforms the edges.
Dub4 / EBN (cont)
● Localisation:
○ Two stage process: place recognition using FABMAP
○ Local-scale localisation generates 6-DoF transform with the
most inliers
● Path memory
○ Resource requirements reduced using a ranking policy based
on the distance between candidate experience nodes and the
current VO-based position estimate
● Illumination invariance
Dub4 Performance
Dub4 Performance
Dub4 in a Space context
● Performs well in the current configuration
● Optimise for planetary surface environments:
○ Improve illumination and viewpoint
invariance
○ Feature detectors optimised for low-featured
environments
Single Camera Navigation
● Wide Angle monocular camera pointing down
● Track ground texture in front of rover
● Nodes store features - edges represent transforms
between nodes (like Dub4)
● Approximately planar assumption - derive planar
induced homography
● Outlier rejection with RANSAC
Single Camera Navigation (cont)
● Used SURF here but others are possible, SIFT, BRIEF
etc
● Recover motion from homography using SVD - reject
unfeasible solutions
● Maintains regular localisation (no dropouts)
● Multiple experiences - drift between runs but can
localise and repeat the path
Mononav Performance
Dense Reconstructions
Dense Reconstructions
Dense Reconstructions
Dense Reconstructions
● Feature rich, lifelike 3-D
meshes
● Tool for mission scientists
● Machine Learning
● Advanced Autonomy
Conclusions and Next Steps
● Oxford VO performance
● Dub4 performance
○ potential to enhance with minor modifications
● Single camera localisation performance
○ minor modifications only
● Dense Reconstructions
○ Science visualisation tool
○ Enhanced AI
○ Robot geologist