The document presents a multiobjective optimization algorithm called MOSA for generating optimal speed profiles for autonomous vehicles. MOSA considers both travel time and energy use as cost functions to be optimized, while ensuring safety constraints are met. It works by chunking a route into segments, generating parametric speed profiles for each chunk based on an intelligent driver model, and using simulated annealing to optimize speed parameters to minimize costs. Tests show MOSA converges quickly and is less sensitive to initial parameter values than other methods. Comparisons to experimental driver data found MOSA outperforms experiments in energy use and respects safety rules better than other algorithms.
Safe and ecological speed profile planning algorithm for autonomous vehicles using a parametric multiobjective optimization procedure
1. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Institut français
des sciences et technologies
des transports, de l’aménagement
et des réseaux
SAFE AND ECOLOGICAL SPEED PROFILE PLANNING
ALGORITHM FOR AUTONOMOUS VEHICLES USING A
PARAMETRIC MULTIOBJECTIVE OPTIMIZATION
PROCEDURE
Olivier Orfila
Dominique Gruyer
Karima Hamdi
Sébastien Glaser
COSYS-LIVIC
2. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Context: Energy and safety
• Energy:
• In France: passenger
cars represents 55% of
CO2 emissions
• Regulations are stricts
for all countries
• Safety:
• In France, 3469 people
died on roads in 2016
• In EU, accidents targets
will be difficult to reach
3. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Objective
• Autonomous driving = miracle solution?
• R&D on autonomous driving are
concentrated on making it real (faisability)
• Objective:
• Develop an algorithm generating optimal
speed profiles for automated vehicles:
• Safety taken as constraints
• Travel time and energy use as cost functions
• Test it by comparing results to actual data
4. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Methodology
• Proposing an algorithm managing several
objectives:
• MOSA (Multiobjective Optimization planning
based on Simulated Annealing)
• Testing its convergence and sensitivity to
several parameters
• Comparing it to experimental results on
open road
5. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Proposed method properties
• Chunked
• Four profiles are
defined
• Multiobjective
• Linear scalarisation
• Optimal
• Simulated annealing
• Parametric
• Combination of
Intelligent Driver
model segments
6. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Algorithm description
• From an origin to a destination,
extracting road data
• From OSM (Open Street Map),
curvature, slope, speed limit,
junctions
• Generating safety constraints on
speed limits, safe curve speed and
junctions
• Chunking speed profile
• Classifying each chunk
• Optimizing speed on each chunk
• Simulated annealing
(*Kirkpatrick,1983) on a parametric
speed profile
• Each solution is a 5 elements vector
describing a speed profile solution
*Kirkpatrick, S.; Gelatt Jr, C. D.; Vecchi, M. P.
"Optimization by Simulated Annealing". Science.
220 (4598): 671–680. (1983)
7. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Chunking
• Constraints
translated into
square signal
• Each chunk is
identified
depending on
raising and falling
edges
8. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Generating a chunk speed profile
• Each chunk is associated to a
speed profile parametric
function depending on its type
• From IDM* (Intelligent Driver
Model) description of
acceleration:
• a: Initial acceleration (ms-2)
• Vdes: desired speed (ms-1)
• pa: percentage of distance of
acceleration phase
• pd: percentage of distance of
deceleration phase
• d: Final deceleration (ms-2)
• Speed optimization using
Simulated Annealing on these 5
parameters
Speed limit
Distance
Speed
Speed profile
A profile
Vdes
pa pd
*Treiber, Martin; Hennecke, Ansgar; Helbing, Dirk (2000), "Congested traffic states in empirical observations and
microscopic simulations", Physical Review E, 62 (2): 1805–1824
Speed limit
Distance
Speed
Speed profileB profile
Vdes
pa pd
9. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Simulated Annealing (SA) algorithm tests
• Convergence
• Before 500 of SA iterations
• Sensitivity
• High to initial values of SA
parameters
• Low to initial values of variables
10. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Selecting a multiobjective
optimal solution
• Challenge:
• Find an optimal solution
with two or more
competing objectives
• A solution:
• Using preference based
(linear scalairisation)
combining two objective
functions (*Deb, 2002)
• Objectives:
• Travel time
• Fuel use
min{𝑓1 𝑥 , 𝑓2 𝑥 }
𝑥 ∈ Ω,
𝑓0(𝑥) = 𝛼1 𝑓1 𝑥 + 𝛼2 𝑓2 𝑥
𝑑𝐸𝑡ℎ𝑒𝑜
=
1
2
𝜌 𝑎𝑖𝑟 𝑆𝐶 𝑥 𝑣2 + 𝐶𝑟𝑟 𝑚𝑔 + 𝑚𝑝 + 𝑚𝑎 𝑣𝑑𝑡,
η =
𝐸𝑡ℎ𝑒𝑜
𝐸 𝑚𝑒𝑎𝑛
= 105
1
2 𝜌 𝑎𝑖𝑟 𝑆𝐶 𝑥 𝑣2 + 𝐶𝑟𝑟 𝑚𝑔 + 𝑚𝑝
𝑓(𝑣)𝑒 𝑐𝑎𝑟𝑏 𝜌𝑐𝑎𝑟𝑏
,
*Kalyanmoy Deb, Multi-objective optimization using evolutionnary algorithms, Wiley, ISBN 0-471-87339-X, (2002)
11. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Comparing to test site experiments:
Description of dataset
• Experimental setup
• A panel of drivers drove
twice on the same road
with the same car
(normal and ecodriving
conditions)
• Test route
• Test vehicle
• Test participants
• 21 drivers (40% female
drivers)
• Ecodriving advice
provided orally before
test drive
12. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Results
• Global results
• Several speed profiles
generated
• They repect safety rules
(curve and limit speeds)
• Pareto plot
• MOSA better than all
experiments on energy
use
• MOSA performing better
than Dijkstra
• MOSA cannot reach
some points (travel time)
13. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Conclusions
• Conclusions:
• A multiobjective speed profile optimization algorithm has
been proposed, tested and compared to experiments
• The proposed method (MOSA) seems to outperform
experimental results and previous method in given
conditions
• Perspectives:
• MOSA is being implemented in electric autonomous
vehicle (with VEDECOM)
• MOSA can be implemented onboard as assistance system
for curve warning and energy efficient speed advice
(ecodriving)
• Need to be extended to dynamic obstacles (reactive
system)
14. Institut français des sciences et technologies des transports, de l’aménagement et des réseaux
www.ifsttar.fr
Thanks for your attention
Olivier ORFILA
Researcher IFSTTAR
Deputy director LIVIC
olivier.orfila@ifsttar.fr
www.olivierorfila.fr
Tél. +33 (0)1 30 84 40 25
LIVIC - Laboratoire sur les Interactions Véhicules-Infrastructure-Conducteurs
25 allée des Marronniers
78000 Versailles
FRANCE