Challenges for Navigation on Mars CL24_6807.pdf

Challenges and solutions for early morning terrain relative navigation on mars.

1. Challenges and Solutions for Early Morning
Terrain Relative Navigation on Mars
Andrew E. Johnson , Nikolas Trawny, Timothy P. Setterfield, Yang Cheng, Jeremy Nash,
Daniel Clouse, Gabrielle Massone, and Miguel San Martin
Autonomous Systems Division
January 9
th
, 2025
Presentation to the AIAA SciTech Forun 2025
This research was carried out at the Jet Propulsion Laboratory, California Institute of Technology, under a contract with the National Aeronautics and Space
Administration. This work was funded by the Mars Sample Retrieval Lander Project in the NASA Science Mission Directorate. The decision to implement Mars
Sample Return will not be finalized until NASA’s completion of the National Environmental Policy Act (NEPA) process. This document is being made available
for information purposes only.

2. Problem Statement
2
•LVS is an implementation of Terrain Relative
Navigation (TRN) that uses images and IMU
data to estimate the spacecraft navigation
state
–M2020 achieved <5m touchdown horizontal
position error
•An Enhanced Lander Vision System is proposed
for Sample Return Lander (SRL)
•Launching SRL in 2031 or 2033 would require a
landing in the early morning
–this is outside the tested operational envelope of
LVS
•Process terrestrial field test and orbital Mars
imagery to quantify the sensitivity of landmark
matching to sun illumination differences
between the reference map and descent
images.
•Develop two novel methods for predicting the
appearance of an early morning reference map
from the available afternoon orbital imagery to
increase the number of correct landmark
matches
M2020 LVS EDL
LVS flight hardware

3. Mars Illumination Conditions
3
CTX 3pm
Map Images
Mars 2020
and 2028
Landing
Images
2031
Morning
Landing
Images
The Issue

4. Field Test Illumination Conditions
4
Terrestrial “Landing”
Test Data
Terrestrial 10am
Map Images

5. December 2023 Field Test
5
•used a Photon Focus camera
•same model as 2014 LVS Field
Testing
•1024x1024 pixels like LCAM
•100˚ FOV > 90˚ LCAM, but close
enough
•Applanix INS/GPS for ground truth
position and attitude
Same Terrain Imaged Under
Different Illumination Conditions
Salton Sea, Southern California
Primary Test Site
(Terrain Relief, Cliffs)
Secondary Test Site
(Bland Terrain)
Flight Paths
Successful Feature Matches (Green)
Despite Large Illumination Difference

6. •Anza Map
•“Depression” target
•NW flight direction
•Morning to afternoon
•Movie shows the images projected
onto the DEM
•Some differences in brightness due to
manual setting of exposure time
•Notice the shadows moving
6

7. Field Test Performance
7
Cliff Target Depression Target

8. CaSSIS Mars Imagery
8
•CaSSIS is an ESA-operated camera on Trace Gas Orbiter (TGO)
•Semi-frame camera, not push-broom
•Not in sun-synchronous orbit, so captures many times of day
•Resolution is 4-5m/px (higher than CTX 6m/px)
•4 bands, used PANchromatic
•Blue: Images Selected for Analysis

9. CaSSIS Illumination Conditions
9
Mars ”Landing”
Test Images
CTX 3pm
Map Images

10. afternoon image
has plenty of fine
landmarks inliers
decrease in fine
landmark inliers
in the early
morning
significant decrease outside of
operational envelope
what happens
in this gap?
coarse landmarks
stay high for afternoon
and morning
decrease in coarse landmarks
for very early morning
approximate SRL landing time
interval and illumination conditions
CaSSIS Performance against Mars 2020 Map (Heritage)
10
•Coarse landmark matching
works well except at sun
elevation angle that are
lower than operation
envelope (25˚)
•Fine landmark matching
shows a definite decrease
with as the illumination
conditions change to
morning
•There is a gap in the
CaSSIS coverage right in
the early morning time, but
there are images on both
sides
•Wide range of inliers
indicates that performance
is dependent on terrain and
not just illumination
conditions

11. Lighting Invariant Matching Algorithm
(LIMA)
11
CTX1 CTX2 CTX3
LIMA LIMA+ PSRM CaSSIS
•Lambertian reflectance
assumption
•LIMA+ uses three images
to compute image A that is
the product of slope and
albedo
–images are radiometrically
corrected
–accurate DEM not required
–independent from lighting
conditions
•Use A to predict image with
early morning lighting
–PRSM = Predicted Surface
Reflectance Map

12. Shading-based Albedo and Normal Estimation
(SANE)
12
CaSSIS Image 005108 (Early Morning) CTX Image (Afternoon)
SANE-AO3 (Noon Image) SANE-RAD-005108 (Early Morning Render)
•Lambertian reflectance
assumption
•SANE uses all images
available to compute an
albedo a weighted surface
normal at each pixel
–images are radiometrically
corrected
–accurate DEM not required
–independent from lighting
conditions
•Use a to predict 3 different
images with early morning
lighting for comparison
–SANE-AO3: PRSM with
overhead illumination
–SANE-RAD: PRSM with early
morning illumination
–SANE-BLN: 50/50 blend of
SANE-AO3 and SANE-RAD
(partially accounts for diffuse
illumination)

13. PRSM Performance with CaSSIS Image
13

14. Landmark Matching Performance for All CaSSIS Test
Images
14
dramatically increase in the
number of fine match inliers
with LIMA and SANE-BLN
one specific image is
challenging for coarse
matching

15. Fine Position Estimation Performance for
All CaSSIS Test Images
15
This westward bias (-X) is likely co-registration error
of CaSSIS images (on-going work)
Even with the bias, the 95%tile residuals are less
than the 30m requirement on horizontal position

16. Comparison of PRSM by Number of
Outliers
16
•Maximum matching error of inlier features ingested by estimator (in meters), as a function of
total number of candidate feature matches going into geometric outlier rejection and their outlier
fraction.
•Mars 2020 EDL feature matches are marked in blue.
•early morning landing with the afternoon M2020 flight map results in unacceptably high outlier
fractions,
•feature count and outlier fractions computed with LIMA6 and SANE-BLN PSRMs yield acceptable
matching performance with margin.
100 RANSAC
iterations
1000 RANSAC
iterations
100 RANSAC iterations limited
to accept only images with at
least 20 estimated inlier
matches.

17. Conclusions
17
•The Mars 2020 LVS system was not designed to handle large changes in
the illumination between the reference map and the descent images used
for TRN.
•Testing with terrestrial field test images and orbital Mars imagery has
shown that landmark matching is sensitive to large changes in sun
azimuth and can fail completely when the scene has significant terrain
relief.
•Two independent methods (LIMA and SANE) were developed that
generate predictions of the reference map for the early morning
illumination conditions expected for a 2031 or 2033 launch of SRL.
•The methods clearly boost the number of correct landmarks up
to a level where position estimation is be reliable and launching
SRL in these windows is feasible.
•Future work is needed to reduce small biases in the estimated positions.