AIMING FOR THE MOON Machine Learning Research for Team Astrobotics - PowerPoint PPT Presentation

AIMING FOR THE MOON Machine Learning Research for Team Astrobotic’s Lunar Lander Research conducted at CMU in conjunction with

Research Behind Lunar Lander Challenges of a Lunar Lander ML Research for the Lander - FTW Lander Perception / Visual Registering Upcoming Work for Dec. 2012 Launch

Motivation of Lunar Exploits  Lunar Resources  Unfiltered solar energy  Clean fusion (helium3)  Methane, ammonia, minerals  Lunar docking bay  No gravity  Available fuel  Watchtower  Base for infrared telescopes (icy craters)

Research Behind Lunar Lander Challenges of a Lunar Lander ML Research for the Lander - FTW Lander Perception / Visual Registering Upcoming Work for Dec. 2012 Launch

It has been done  1969 – Apollo 11  manned mission  1970 – Luna 17 with Луноход (Lunokhod 1)  remote controlled robot View from Camera 2 Images from Wikipedia and www.mentallandscape.com

So where is the novelty?  Pinpoint Landing - 500 m  Less than a hundredth of a degree (lat/long)  High autonomy  Little or no human support in landing  Preset trajectory  Fault detection  Error recovery  Proof of concept – reliable commercial lander

Stages of Landing  Orbit insertion  Determine Lander orientation

Stages of Landing  Orbit insertion  Determine Lander orientation  De-orbit and breaking  Determine orbital parameters

Stages of Landing  Orbit insertion  Determine Lander orientation  De-orbit and breaking  Determine orbital parameters  Descent – 18km to 500m  Keep track of Lander coordinates

Stages of Landing  Orbit insertion  Determine Lander orientation  De-orbit and breaking  Determine orbital parameters  Descent – 18km to 500m  Keep track of Lander coordinates  Touch down  500m to 150m: compute slopes, detect craters  Less than 150m to surface: detect small obstacles

Challenges – Perception  Detect whether Lander is off course  Detect whether sensors function properly  NO CAMERA - for a while  In case of off-course landing, pick landing spot

Research Behind Lunar Lander ML Research for the Lander - FTW Lander Perception / Visual Registering Upcoming Work for Dec. 2012 Launch

Research - Vision  Landscape identification  Feature tracking  Mega-structures  Craters  Obstacle detection

Research – Evidence Fusion  Density Elevation Map (DEM) construction  Sparse LIDAR readings  Images of surface  Combine sensor readings to obtain position  Voting scheme to determine faulty components

Research – Knowledge and AI  Decision unit in the Lander  Inference to determine  Position  Lander State  Planning  Save fuel resources  Pick landing spot to facilitate rover movement

Research Behind Lunar Lander Lander Perception / Visual Registering Upcoming Work for Dec. 2012 Launch

Lander Sensor Array Stage Requirement Sensor Orbit insertion Attitude IMU Sun position Sun tracker/Star tracker Pre-defined Orbital Parameters Deorbit Attitude Camera/IMU Descent stage Attitude Camera/IMU (18kms to 500mts) Planning Touch down Attitude Camera/IMU (500mts to ground) Altitude Pointed RADAR Slope of Ground LIDAR Velocity Doppler Surface characteristics LIDAR/Camera

Visual Registering  Step 1: Crater Detection Crater Detection on LRO Image Crater Detection on Image captured by camera on the Black Magic platform

Visual Registering  Step 2: Comparing Landscape to Stored Data Test Set True False True False Positives Positives Negatives Negatives Apollo 11 4/5 3/20 17/20 1/5 Apollo 14 3/5 4/20 16/20 2/5 Apollo 16 5/5 3/20 17/20 0/5 Apollo 17 4/5 4/20 16/20 1/5

Digital Elevation Map Sparse LIDAR data (Lunar Reconnaissance Orbiter

Digital Elevation Map Image (LCROSS Impact)

Markov Random Field X – image Z – sparse elevation measurements Y – estimated elevation map L – points where elevation readings exist N(i) – neighborhood of point i on the grid w ij – correlation between pixels in image Z X LIDAR (1% of available data) Image of cabeus crater

MRF with Shading Coefficient X – image Z – sparse elevation measurements Overlap of Y – estimated elevation map image and L – points where elevation readings exist LIDAR N(i) – neighborhood of point i on the grid w ij – correlation between pixels in image Pixel correlation Shading coefficient p =5% quantile of image data – the shaded pixels

Experiment - Terrain Model

Experiment – Model + LIDAR

MRF Results Elevation map after interpolation • mean error 175.20 m • 16.09% of average elevation Elevation map after 200 iterations of Coordinate Descent • mean error 143.75 m • 13.2% of average elevation

Scanning for a landing site  Image-based landing site selection  Elevation-based landing site selection

Research Behind Lunar Lander Upcoming Work for Dec. 2012 Launch

Work for December 2012 Launch  Feature Tracking  AI unit  Decisions  Inference  Planning  Integrated Testing Framework