Stanley

1
Stanley

• The Robot That Won The DARPA Grand Challenge

2
The 2005 DARPA Grand Challenge

• Desert race designed to test the capabilities of
  autonomous robot design
• Spanned 175 miles of rough terrain
• Goal of the challenge is to complete the course
  in under 10 hrs
• Grand Prize 2M
• Race Day Oct 8th 2005
• 195 teams registered.23 raced (of which only 5
  finished)

3
DARPA Grand Challenge cont.

• Robots all competed on the same course
• Started one after another at 5 minute intervals
• If passing occurred, DARPA officials would pause
  the lead vehicle to allow passing(this allowed
  robots to deal with other cars as static
  obstacles)

4
Waypoint Definitions

• Race officials provide complete race description
  in RDDF format.
• Contains waypoints with
• Latitudes
• Longitudes
• Corridor Widths
• Associated Speed Limits
• 2005 Challenge contained 2935 waypoints
• Contestant robots are expected to navigate the
  course solely via this data and their own
  percepts

5
Stanley

• The product of Stanford Universitys autonomous
  driving research
• Was the grand prize winner of the 2005 race
• Placed first with a race time of 6 Hrs, 53 min,
  58 seconds
• Based off of the 2004 Volkswagen Touareg R5 TDI

6
More Stanley

• Equipped with
• 5 Laser Finders (25m range)
• 2 Radar Sensors (200m range)
• Three Antennae(1 for GPS Positioning, 2 for GPS
  Compass)
• DARPA Wireless E-Stop System
• 1 Color Video Camera (70m range)
• Computer systems
• 6 Pentium M CPUs
• 3 for running race software
• 1 for logging race data
• (2 idle)
• 1 Gb Ethernet Switch
• All systems run Linux

7
Race Day

• AIPPT\RaceDay.wmv

8
Stanley Software System
9
Design Philosophy

• Modular approach (approx. 30 run in parallel)
• 1. Sensor Interface Layer
• Responsible for receiving sensor data directly
  from devices
• 2. Perception Layer
• Maps sensor data to internal modules
• Three different mapping modules construct 2D
  environment maps based on raw data from sensor
  interface layer
• 3. Control Layer
• Responsible for regulating physical actuators
  (brakes, steering, gas, etc)
• State represented by a Finite State Machine (FSM)
• 4. Vehicle Interface Layer
• Responsible for interfacing with the various
  physical components of the vehicle as well as the
  main power server.
• 5. User Interface Layer
• Comprised of an emergency override switch and an
  in-cab debugging panel
• 6. Global Services Layer
• Provides basic services like device naming and
  communication, module health monitoring, power
  supply and regulation, and individual device
  state management, and device clock
  synchronization.

10
Laser Mapping System

• Each laser in the array (5) generates 181 range
  measurements (up to 25m) spaced 0.5º apart and
  polled at 100hz
• These measurements are then converted to (X, Y,
  Z) coords by the Laser Mapper on-the-fly
• Obstacles are detected by the Laser Mapper if we
  can find two nearby points whose vertical
  distance Z1 Z2 exceeds some critical value d
  (initially 15cm)
• The terrain is considered to be drivable if no
  such set of points are found
• False positives can be triggered if disturbances
  in the vehicles orientation (pitch, yaw, roll)
  occur.
• In practice, using the vertical distance analysis
  yields a false positive rate of 12.6

11
Laser Mapping System (contd.)

• To avoid making false positives, we employ
• Probabilistic analysis
• A discriminatory algorithm
• Boils down to
• If p(Z1 Z2 gt d) gt a,
• Perform on-the-fly adjustment of d
• Future readings will take updated value into
  account when marking terrain/objects
• Result
• Drops original false positive rate of 12.6 to
  0.002
• This is a VAST improvement over the vanilla test
  case.

12
Visual Terrain Analysis

• Long-range obstacle detection (up to 70 meters)
• Supports speeds just above 35 MPH
• Processed in 4 stages
• (a) Raw image data
• (b) A processed image with pixel classification
  and a laser quadrilateral
• (c) Pixel classification prior to thresholding
• (d) Horizon detection for removing the sky
• Includes visual adaptation to adjust for dynamic
  road and light conditions

13
Road Boundaries

• Road boundaries are calculated locally via laser
  map data
• Stanley avoids the majority of road boundaries by
  following the road along its center
• Boundaries are calculated as parallel beams of
  which Stanleys path converges towards the center
• Sampling in z accelerometer data provides shock
  value, from which Stanley can determine
  appropriate velocities

14
Path Planning

• In order for Stanley to reliably use global path
  planning the RDDF data is initially smoothed
• Stanleys coordinate system is based off of the
  calculated route, where lateral offset is the
  perpendicular distance to the fixed base
  trajectory
• From Stanleys perspective altering the lateral
  offset correlates to moving left or right from
  the projected path
• The base trajectory is a smoothed version of the
  provided RDDF
• By implementing smoother turns Stanley may
  traverse turns more sharply and more safely
• Since the RDDF contains a limited number of
  waypoints the curvature of the terrain is poorly
  predicted
• By smoothing natural curvature we may compute a
  more direct path

15
Path Planning (cont.)

• Initial path computation is 1 hour prior to the
  event (upon the receipt of the RDDF)
• The base trajectory computation is performed in a
  4-stage procedure
• 1. Points are added to the RDDF to account for
  local curvature
• 2. Coordinates of all the upsampled points are
  adjusted to minimize the curvature of the path
• 3. Cubic spline interpolation is then applied to
  obtain a path that is differentiable
• 4. Speed limits from the RDDF, from lateral
  acceleration constraints, and from bounded
  deceleration are then added into the base
  trajectory
• The base trajectory computation is then saved to
  a different file than the original RDDF to allow
  Stanley to still check bounds required by DARPA

16
Velocity Management

• 3 Components submit velocity suggestions into the
  velocity controller
• Planner-Provides legal speed limits per the RDDF
• Velocity Recommender- Provides a linear
  acceleration and a quick deceleration
  additionally limiting velocity from slope
  adjustments
• Health Monitor- Suggests velocities based on
  vertical acceleration caused from rough terrain
• Once this data is input the minimum is chosen and
  implemented by the velocity controller
• Pending the result of the velocity computation
  the car chooses to brake or provide throttle as
  calculated
• Steering control is then generated based on the
  path planner, pose/velocity estimates, and the
  measured steering angle

17
Results

• After arriving for race day the RADAR system was
  disconnected
• The RADAR system was deemed redundant as the
  vision-based percept detected long-range
  obstacles as well
• The USB driver of the system was causing trouble
• After the qualifying runs Stanley was deemed the
  second starting position
• Stanley was paused twice, one of which was to
  allow CMUs H1ghlander to pass them
• H1ghlander was later passed by Stanley
• Stanley finished with a time of 6 hrs, 53 min, 58
  seconds with an average velocity of 19.1 mph

18
Other Stats

• Stanleys velocity was derived from
• 68.2 Pre-calculated (DARPA speed limits and
  lateral acceleration constraints in turns)
• 18.1 Rugged or steep terrain encountered
• 13.1 Vision module (limits speed to 25 mph)
• .6 GPS outage
• Total of 17 incidents of laser stream stalls
  (most of which occurred between Mile 22 and Mil
  35)
• This caused 4 incidents of phantom object
  detection, resulting in a major swerve
• Stanleys average lateral offset was 74 cm

19
DARPA Grand Challenge 2007

• 60-mile urban area driving course
• Involves obeying traffic laws while merging into
  moving traffic, navigating traffic circles,
  negotiating busy intersections, and avoiding
  obstacles
• Required finish time of under 6 hours
• First place reward has been changed to just a
  Trophy

20
Questions?

• Works Cited
• Stanely The Robot That Won The DARPA
  Challenge. Journal of Field Robotics 23(9)
  (2006) 661-692.
• Walker, Jan. DARPA Announces Third Grand
  Challenge. October 2, 2006.