Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning...
Transcript of Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning...
![Page 1: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/1.jpg)
10/16/2008
1
Self-Supervised Online Learning Approaches for Robot Navigation
16-831, Fall 2008
October 16
2
Mobile Robot Navigation
![Page 2: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/2.jpg)
10/16/2008
2
3
Local Perception System
Local
Features
Perception systemSensors
Planner (D*)
Cost map
4
Motivation
Sensing range
Onboard perception system loses effectiveness at longer ranges (past 12-15 meters in this case)
Results in inefficient and often dangerous exploration
![Page 3: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/3.jpg)
10/16/2008
3
5
Motivation
Sensing range
Onboard perception system loses effectiveness at longer ranges (past 12-15 meters in this case)
Results in inefficient and often dangerous exploration
6
How to Improve
Use overhead data (imagery, elevation, etc.)
![Page 4: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/4.jpg)
10/16/2008
4
7
Hand-Train Overhead Interpreter
Hand-train overhead classifier / cost predictor
Apply to larger map
Use resulting map for planning
8
How to Improve
Use overhead data (imagery, elevation, etc.) Difficult to interpret consistently
Variations in terrain, lighting, weather, time of gathering
Extend the range of the perception system Not enough data to accurately
generate perception system’s features Can’t estimate ground plane,
inaccurate density, etc.
Features that are computableare difficult to interpret consistently
![Page 5: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/5.jpg)
10/16/2008
5
9
Overhead data features
Color, texture, clustering, elevation, PCA, neighbor features, etc.
Far-range sensor data features
Color, ladar point elevation spread / std, neighbor features, etc.
How can we best use these potentially powerful, but difficult to generalize, features?
How to Improve
Scoped Learning
10
![Page 6: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/6.jpg)
10/16/2008
6
11
The Algorithm
Model the relationship between these features and measured traversal cost in a Bayesian probabilistic framework
Online Bayesian Linear Regression
Relate multiple data sources with different “scope” to each other
Blei, ’02
12
The Algorithm
Initialize the distribution to the prior distribution
We want to compute
For every training example , multiply the distribution by
By computing , we are performing self-supervised learning using a Bayesian linear regression model
![Page 7: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/7.jpg)
10/16/2008
7
13
The Algorithm
Learn to interpret these locale-specific features by taking advantage of the globally interpretable features from the perception system
14
The Algorithm
Learn to interpret these locale-specific features by taking advantage of the globally interpretable features from the perception system
![Page 8: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/8.jpg)
10/16/2008
8
15
The Algorithm
Learn to interpret these locale-specific features by taking advantage of the globally interpretable features from the perception system
16
The Algorithm
Learn to interpret these locale-specific features by taking advantage of the globally interpretable features from the perception system
![Page 9: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/9.jpg)
10/16/2008
9
17
The Algorithm
Learn to interpret these locale-specific features by taking advantage of the globally interpretable features from the perception system
18
The Algorithm
Learn to interpret these locale-specific features by taking advantage of the globally interpretable features from the perception system
![Page 10: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/10.jpg)
10/16/2008
10
19
Results
Overhead Online Learning
Online use
Offline use
Far-Range Online Learning
20
Results
Overhead Online Learning
Online use
Offline use
Far-Range Online Learning
![Page 11: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/11.jpg)
10/16/2008
11
21
Results
Overhead Online Learning
Online use
Offline use
Far-Range Online Learning
22
Overhead Online LearningUsing features from 40cm color imagery and elevation dataUpdating traversal cost map onboard robot in 65 m radius
![Page 12: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/12.jpg)
10/16/2008
12
23
Overhead Online Learning
24
Results
Overhead Online Learning
Online use
Offline use
Far-Range Online Learning
![Page 13: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/13.jpg)
10/16/2008
13
25
Far-Range Online Learning
26
Far-Range Online Learning
![Page 14: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/14.jpg)
10/16/2008
14
27
Results
Far Range Online Learning
28
ResultsOverhead
Online Learning
![Page 15: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/15.jpg)
10/16/2008
15
29
Results
Overhead Online Learning
Online use
Offline use
Far-Range Online Learning
30
Overhead Online Learning (Offline)
Training course
2000m x 750m
![Page 16: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/16.jpg)
10/16/2008
16
31
Overhead Online Learning (Offline)
Using 1m black and white imagery data
Using 35cm color imagery data
32
Overhead Online Learning (Offline)
Data alignment Use to
detect most likely map alignment
Use alignment with the highest average log probability over all examples seen
![Page 17: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/17.jpg)
10/16/2008
17
33
Additional Benefits
Reversible learning
Confidence-rated predictions
34
Reversible Learning
Multiple estimates of single quantity
Receive example
Receive lower variance estimate
always takes into account only best estimates available for all examples
![Page 18: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/18.jpg)
10/16/2008
18
35
Additional Benefits
Reversible learning
Confidence-rated predictions
36
Confidence-rated predictions
Use variance estimate (HW3!) provided by algorithm for the probability of each estimate as measure of confidence
“Surprise” at seeing set of features
![Page 19: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/19.jpg)
10/16/2008
19
Far-Range Online Learning with Velodine
37
Movie…
Questions?
38
![Page 20: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/20.jpg)
10/16/2008
20
A Self-Supervised Terrain Roughness Estimator for Off-Road Autonomous Driving
David Stavens and Sebastian Thrun
Stanford Artificial Intelligence Lab
“Combines” strengths of multiple sensors.
Ultra-Precise, No Range Precise, Long Range
![Page 21: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/21.jpg)
10/16/2008
21
Velocity Planning for DGC 2005
Mobile robotics traditionally focuses on steering.
But speed is also important.
Beyond stopping distance and lateral maneuverability.
Stanley adapted its speed to terrain conditions, minimizing shock:
Increases electrical and mechanical reliability.
Mitigates pose error for laser projection.
Increases traction for improved maneuvers.
Correlated with slowing on “hard” terrain.
Simple three state algorithm:
Drive at speed limit until shock threshold exceeded.
Slow to bring the vehicle within the shock threshold.
Uses approx. linear relationship between shock and speed.
Accelerate back to the speed limit.
Reactive Approach (used during DGC)
![Page 22: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/22.jpg)
10/16/2008
22
Acquiring a 3D Point Cloud
Movie…
Errors in Pose and Projection
Goal: know amount of error that is expected so that actual rough terrain or obstacles may be better identified.
![Page 23: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/23.jpg)
10/16/2008
23
Z Error vs. Time
More than t
“Spread” of plot implies more factors than t.
Also related to:
Amount/rate of pitching.
Distance between the two scans.
![Page 24: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/24.jpg)
10/16/2008
24
Comparing Two Laser Points
Uncertainty = pair =
1| z | 2 –
3| t | 4 –
5| xy distance | 6 –
7| dpitch1 | 8 – 7| dpitch2 | 8 –
9| droll1 | 10 – 9| droll2 | 10
Seven Features: z, t, xy distance, dpitches, drolls
10 Parameters: 1 2 … 10 (generated with self-
supervised learning)
Estimate Roughness
Combine points in estimated future locations of wheels to estimate a roughness score, R, for terrain patch.
But how do we assign target values to R?
![Page 25: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/25.jpg)
10/16/2008
25
Self-Supervised Learning
Actual shock when driving over terrain modifies belief about original laser scan.
Improves classifier for subsequent scans!
Self-Supervised Learning
Actual shock when driving over terrain modifies belief about original laser scan.
Improves classifier for subsequent scans!
![Page 26: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/26.jpg)
10/16/2008
26
Caveat: Must Correct for Speed
Mapping from R to Shock
Learn a simple suspension model in parallel with the classifier:
Rcombined = Rleft 11
+ Rright 11
Rleft and Rright is for the terrain under each wheel.
![Page 27: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/27.jpg)
10/16/2008
27
Learning Parameters
Tp = true positive rate
Fp = false positive rate
Maximize Tp – λFp
Used λ = 5 to minimize false positives
Optimized through coordinate ascent
Greedily optimize each parameter individually, decreasing learning rate each cycle by factor of 2.
53
![Page 28: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/28.jpg)
10/16/2008
28
Self-Supervised Monocular Road Detection in Desert Terrain
Hendrik Dahlkamp, Adrian Kaehler, David Stavens, Sebatian Thrun, and Gary Bradski
Stanford University, Intel Corporation
![Page 29: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/29.jpg)
10/16/2008
29
Goal: Detect drivable surface for aiding speed calculations
57
Extract “training” area using laser data
58
Project onto camera image
Assume that area contains only drivable surface
Remove sky and shadows
Range: ~22m
![Page 30: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/30.jpg)
10/16/2008
30
Learn visual model of nearby road
Approximate using mixture of k Gaussians in RGB space
Additional Gaussians describing training history
59
Score visual field by road model
Use distance from each pixel to nearest Gaussian to assign a “roadness” score.
60
![Page 31: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/31.jpg)
10/16/2008
31
Select identified patches
Threshold image points further away than 3σ to get a binary drivability image.
Run several filters to remove small non-drivable areas (rocks, leaves) while preserving bigger obstacles.
61
Usage
Used as pre-warning system for capping speed (if can’t see clear road for 40m).
Ran at 12fps on single processor on 320 x 240 images.
Extended road detection to up to 70m.
62
![Page 32: Improving Robot Navigation through Self-Supervised Online ... · Self-Supervised Online Learning Approaches for Robot Navigation 16-831, Fall 2008 October 16 2 ... probabilistic framework](https://reader035.fdocuments.us/reader035/viewer/2022062922/5f0935b57e708231d425beb8/html5/thumbnails/32.jpg)
10/16/2008
32
Pretty video…
63