Professional Documents
Culture Documents
Leonard Tartan Slam
Leonard Tartan Slam
of SLAM
Dr. John J. Leonard
MIT CSAIL and MechE
July 2021
C S A I L M R G
marinerobotics.mit.edu John J. Leonard, MIT
Outline
• My Personal Research Journey: SLAM, AUVs and Self-Driving
• Q&A
Outline
• My Personal Research Journey: SLAM, AUVs and Self-Driving
• Q&A
My Background
“Toward Lifelong Visual Localization and Mapping”, H. Johannsson, PhD Thesis, MIT/WHOI Joint
Program, June 2013 (REMUS AUV data courtesy of Doug Horner, NPS)
Amazing Students
• Q&A
https://global.oup.com/academic/product/how-to-get-your-phd-9780198866923?cc=us&lang=en&
Hugh Durrant-Whyte
Grasp Lab PhD
1986
Rodney Brooks
“These problems are far
from solved.”
“much remains to be done
to produce complete
solutions
Representation: Occupancy Grids
ICRA 1985
Position Referencing and Consistent World 1985
Raja Chatila
Raja Chatila
Consistent Integration and Propagation of
Disparate Sensor Observations
Hugh Durrant-Whyte
O3
O1 O2
observation
R1
R0
motion
1998
Visual SLAM –
Andrew Davison
Feature-based SLAM – Error
Consistency and Data Association
J.-S. Gutmann and K. Konolige. Incremental Mapping of Large Cyclic Environments, in: International
Symposium on Computational Intelligence in Robotics and Automation (CIRA'99), Monterey,
November 1999.
First SLAM Summer School Stockholm, 2002
First SLAM Summer School Stockholm, 2002
TartanSLAM CMU, 2021
TartanSLAM CMU, 2021
Outline
• My Personal Research Journey: SLAM, AUVs and Self-Driving
• Q&A
Recent Research in My Group
O3
O1 O2
observation
R1
R0
motion
Kevin Doherty, MIT
Combining Objects and Locations for Spatial Navigation
building, car, …
aka Objects
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
The hypothesis set grows exponentially
Hypotheses
… and so on
New landmark
Known landmark
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
Our semantic SLAM approach
Landmarks
New landmark
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
Our semantic SLAM approach
Landmarks
Known landmark
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
We implicitly represent multiple hypotheses
1)
Hypotheses
2)
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
We implicitly represent multiple hypotheses
1)
Hypotheses
2)
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
We implicitly represent multiple
hypotheses
1)
Hypotheses
~ 200 AprilTags
~ 30 minutes of data
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
Experiment with AprilTags in MIT
Corridors
MIT RACECAR robot
~ 200 AprilTags
~ 30 minutes of data
“Probabilistic Data Association via Mixture Models for Robust Semantic SLAM,”
K. Doherty, D. Baxter*, E. Schneeweiss*, J. Leonard. In IEEE International Conference on Robotics and Automation 2020.
Ambiguity-Aware Object-
Based Navigation
Ziqi Chad Kevin
Hypothesis
Decisions
Ziqi Lu*, Qiangqiang Huang*, Kevin Doherty, and John J. Leonard. Consensus-Informed
Optimization Over Mixtures for Ambiguity-Aware Object SLAM. To Appear in IROS 2021
https://arxiv.org/abs/2107.09265
Idea for testing ambiguity-aware object mapping during Covid …playing cards
56
Object Pose Ambiguity
Robot fails to perceive the full 6D pose of an object
57
Object Pose Ambiguity
Hypo 2
Hypo 1
58
Long-Term Goal: Ambiguity-Aware Object SLAM
1. Cadena, Cesar, et al. "Past, present, and future of simultaneous localization and mapping: Toward the robust-perception age." IEEE Transactions on robotics 32.6 (2016): 1309-1332.
59
2. Salas-Moreno, Renato F., et al. "Slam++: Simultaneous localisation and mapping at the level of objects." CVPR. 2013.
Object SLAM Back-End: MAP Estimation
66
1. Wong, Jay M., et al. "Segicp: Integrated deep semantic segmentation and pose estimation." IROS. IEEE, 2017.
Foundation: Factor Graph Optimization
Robot poses
Prior factor
Measurement factor
Odometry factor
(Non-Gaussian)
Hypothesis
Decisions
Time 1
Time 2
Time 3
69
Existing Method: Maximum Likelihood Estimation
Maximum Likelihood
Estimation (MLE)
✔
❌
Brittle with noisy or outlier
measurements
❌
✔
❌
✔
70
Existing Method: Multi-Hypothesis Tracking
Multi-Hypothesis Tracking
(MHT)
Incremental optimization +
efficient multi-hypothesis inference
71
1. Hsiao, Ming, and Michael Kaess. "MH-iSAM2: Multi-hypothesis iSAM using Bayes Tree and Hypo-tree." 2019 ICRA. IEEE, 2019.
Existing Method: Gaussian Mixture Models
Max-mixtures
72
1. Olson, Edwin, and Pratik Agarwal. "Inference on networks of mixtures for robust robot mapping." The International Journal of Robotics Research 32.7 (2013): 826-840.
Max-Mixture Factor
Ambiguous Max-mixture
Measurements factors
Max-mixture factor
Ambiguous
objects
Time 1
Time 2
Time 3 75
Landmark Re-Initialization
76
1. Kaess, Michael, et al. "iSAM2: Incremental smoothing and mapping using the Bayes tree." The International Journal of Robotics Research 31.2 (2012): 216-235.
Infer Dominant Pose Hypothesis
Cards
1.5
1
1st card re-detection (JC)
0.5
0
0 100 200 300 400 500 80
t (s)
Summary
Back-end Optimizatio
• Q&A
Discussion
• Robust Semantic SLAM is not just an “engineering” problem
• The library of objects for available for Semantic SLAM is too small
• Q&A