2017/9/24 1 Generalized Predictive Planning for Autonomous Vehicles Scott Pendleton and Marcelo H. Ang Jr. Department of Mechanical Engineering National University of Singapore
2017/9/24 1
Generalized Predictive Planning for Autonomous Vehicles
Scott Pendleton and Marcelo H. Ang Jr.
Department of Mechanical EngineeringNational University of Singapore
2
Why Autonomous Vehicles? (Singapore Perspectives)
• Reduce car ownership– Ride sharing, delivery, logistics
• Efficient use of resources– Car, road infrastructure, less parking spaces
• Public transportation– Last mile/first mile problem– Urban driving as opposed to highways
• Improved Productivity & Safety, “greener”
3
AvailabilityAccessibility
Affordability
Autonomy Ride Sharing
• Multiple vehicle classes: Operational advantages for each vehicle class favor different environments. A combined multi-class service can extend the operational area. True point-to-point service coverage is achievable.
• Disruptive technology: Automation can enable new ways of thinking about automobiles and transportation systems in general. In particular, it can provide affordable, convenient, on-demand mobility.
Autonomous Mobility‐on‐Demand• Vehicle sharing for first-and-last-mile transportation
INTRODUCTION & MOTIVATION
4
Environments
• Road • Pedestrian
5
SMART=NUS Fleet
6
What we can confidently do?• Reactive control with guaranteed safety (lowest layer – always on)
• Mapping and Localization• Local planning
– RRT* variant– POMDP
• Execution & Control– More accurate path following using kinematic constraints
7
Mobility on Demand using Multi‐Class Autonomous Vehicles
8
• One North:– Jan 2015 – 6 km route
– Sept 2016 – 12 km route
– 23 June 2017 – 55 km ‐NUS & Science Pk
8
• 9 vehicles– SMART-NUS: 1– Nutonomy: 6 – Delphi : 1– A*STAR: 1
9
One North – Live Testing
10
Pedestrian crossing Signalized Intersection
Complex intersection Road construction Road construction and jay walking
One North – May 2017
11
Public Deployment at theChinese & Japanese Gardens (Oct 2014)
‐ Long Term Vehicle Testing
‐ To raise awareness‐ To gain public acceptance
6 Days360 km
500 Visitors220 Trips225 Surveys
98% “would ride again”
12
13
our autonomous mobility scooter
14
Our Planning Framework
• Interface planning modules with perception and control modules
• Incorporate acceleration constraints• Establish replanning timing/retriggering• Safety mechanism design for predictive planning
PREDICTIVE PLANNING FRAMEWORK
15
Planning Framework OverviewPREDICTIVE PLANNING FRAMEWORK
16
Planning Framework Overview• Booking System & Mission Planner
• Mobile phone access to webserver for handling mission requests as {Pickup Station, Dropoff Station}
• Dijkstra search over directed graph of reference path segments
• Mapping/Localization• Vertical features extracted from 3D point cloud gathered from 2D LIDAR “rolling window” accumulation over time
• Obstacle Detection• SVM performed over spatio‐temporal features of object clusters from 2D LIDAR
PREDICTIVE PLANNING FRAMEWORK
17
Planning Framework Overview• Cost Map Generator
• Obstacle avoidance cost set for grid locations in a 3D cost map layered by time dimension, up to a time horizon
• Goal Generator• Goal state set at constant distance ahead along route plan
• Steering Control• Pure‐pursuit steering find constant radius arc target to forward waypoint
• Speed Control• Proportional Integral (PI) controller with switching mechanism for throttle vs. braking
PREDICTIVE PLANNING FRAMEWORK
18
Trajectory Planner• Control and Path Guided RRT* (CPG‐RRT*)
– Use RG, path guided sample biasing, and min‐jerk edge connection
PREDICTIVE PLANNING FRAMEWORK
• Same structure of RRT*, but redefine subfunctions:– “Nearest” is RG NN search– “SampleFree” uses biasing– “Line” uses an min‐jerk
profile interpolation along Dubins car paths
– “Steer” and “CollisionFree” are built off the “Line” function
19
Trajectory Planner: SampleFree
PREDICTIVE PLANNING FRAMEWORK
• Retain previous iteration knowledge by Φi‐1
• Bias toward route plan by Φpp
• SampleGoalfor greedy search
• RG Sample for efficient exploration
20
Trajectory Planner: Line
PREDICTIVE PLANNING FRAMEWORK
• Controllable trajectory generation to enforce:– Minimum turning radius (Dubins curves)– Velocity bounds– Acceleration bounds
• Edges are min‐jerk optimal for comfort– Minimizes – Known to be 5th degree polynomial for position
• Trajectory defined over Dubins x Velocity x Time– Configuration space
21
Trajectory Planner: Line
PREDICTIVE PLANNING FRAMEWORK
• First, solve for Dubins curve in SE(2) space• Then, solve for position, velocity, and acceleration w.r.t time by system of equations for boundary conditions:
• Known: pinit , vinit , ainit , pfinal , vfinal . set afinal = 0• Solve for constants b0 … b5
22
Trajectory Planner: Line
PREDICTIVE PLANNING FRAMEWORK
• Polynomial solutions found quickly• Bounds checked over time interval at endpoints and roots
23
Replan Timing
PREDICTIVE PLANNING FRAMEWORK
• Each plans is generated while previous plan is executed
24
Safety Checking
PREDICTIVE PLANNING FRAMEWORK
• Each solution plan is rechecked against an updated observation before execution
• A new variant of braking Inevitable Collision State (ICSb) is applied for passive safety:– A braking maneuver must exist from the commit state following
the solution trajectory to satisfy dynamic minimum braking distance
– Otherwise, velocity profile of solution is overridden by constant deceleration profile up to braking distance
• “Clear zone” applied to command stop when obstacles are very close
25
Control Interfacing
PREDICTIVE PLANNING FRAMEWORK
• Planner must know next commit state as root for plan tree– Control and/or localization error may affect true pose– s1 is expected commit state at end of trajectory Φ0 , but instead
arrive at s1’– Where to begin plan Φ2? Introduce pose correction factor!– Start plan Φ2 from state s2+ w Δs1 (we use w = 0.5)
26
Control Interfacing
PREDICTIVE PLANNING FRAMEWORK
• Pose correction in practice:– Red is odometry trace (series of vectors)– Yellow is commit path– Overlap correlates with velocity undershoot, gap for overshoot
27
Summary: Planning FrameworkPREDICTIVE PLANNING FRAMEWORK
• Predictive planning framework– Real‐time replanning in space‐time
• Trajectory planning algorithm (CPG‐RRT*)– Generates min‐jerk controllable edge connections– Biased sampling for
• Near previous solution trajectory• Near pure pursuit steering trajectory to route plan• Near goal• Reachable configuration space
• Passive safety assurances through adapted braking Inevitable Collison State Avoidance (ICSb)
28
Software Overview
Fleet Management
System Server
Booking App
Multi-Class Autonomous Vehicles
Users
Onboard Verification
VEHICLE PLATFORM DEVELOPMENT
29
Software OverviewVEHICLE PLATFORM DEVELOPMENT
30
Hardware Overview• Common Sensor Suite
• IMU & wheel encoders for odometry• 1 2D LIDAR for Mapping & Localization (M&L) – fuse w/odom
• ≥1 2D LIDAR for Obstacle Detection (OD)• Similar Power Management & Off‐the‐shelf Computers• Ubuntu 14.04, ROS Indigo, i7 processor, 16GB RAM, SSD
• Differing Actuation Mechanisms to Control:• Steering• Braking/Throttle• Gear Selection (Forward/Reverse)
VEHICLE PLATFORM DEVELOPMENT
31
Hardware OverviewStart with a personal mobility scooter, then add…
VEHICLE PLATFORM DEVELOPMENT
32
Hardware OverviewStart with a golf car, then add…
VEHICLE PLATFORM DEVELOPMENT
33
Hardware OverviewStart with a road car, then add…
VEHICLE PLATFORM DEVELOPMENT
34
Safety Overrides• User Button Controls:• Pause• Auto• Manual
• E‐stops, onboard and remote
• Visualizations onboard show perception data and planned path
• Audio cues for station arrival/departure
VEHICLE PLATFORM DEVELOPMENT
35
Experiment Setup• Look for positive emergent behaviors• Compare against baseline planning method:
• Decoupled spatial path and velocity planning• Enlarge obstacle bounds forward based on velocity to treat environment as static
• Trigger replanning only when at a stop due to blockage• Test Scenarios:
• Pedestrian navigation• T‐junction• Defensive driving• Overtaking
VEHICLE PLATFORM DEVELOPMENT
36
Experiment SetupVEHICLE PLATFORM DEVELOPMENT
• Planning visualization
37
• Video available on YouTube: search “FMAutonomy” channel
Predictive Planning Video
https://youtu.be/eVVGZxp03Hc
EXPERIMENTAL VALIDATION
38
• Reactive Control – Guaranteed Safety as a Baseline
• Generalize predictive planning– Plans coupled spatial path and velocity– Demonstrated over varied vehicle types and environments in high‐risk scenarios
• Reachability Guidance– Speed improvement by factor of 9‐10
• Predictive Planning Framework– CPG‐RRT* (biased sampling and min‐jerk edges)– Modified ICSb passive safety assurances
What have we achieved?
39
Towards Mapless Navigation
39
• You are “here” (blue circle)
• Go to #02‐16
• Giving intelligence to robot– To read maps– Navigation to
points in the map
What’s Next?
40
Learning how to drive• Cars and peoplearound
• Movingdirections
• Relativepositions
• Speeds• IntermediateGoal
40
• Steering• Brake• Throttle
What’s Next?
41
Marcelo H ANG [email protected]