RealTime Evolution of Collision Avoidance Behavior using Genetic Programming

1
Real-Time Evolution of Collision Avoidance
Behavior using Genetic Programming

• Matt Luciw
• CSE 848

2
Collision Avoidance

• Basically dont run into anything.
• A local, lower-level behavior, typically used
  with a path-planner.

Use information from local sensors (i.e., not
GPS) to control the agent so that a collision
does not occur. Its better to avoid unnecessary
detours while avoiding the collision (if
possible).
3
Sensor Laser Range-Finder

• The sensor used in the simulation.
• Gives the distance to the nearest point from 0 to
  180 degrees.
• Half-angle resolution, so 361-dimensional input.

4
Tools Lil-GP / CSim

• Lil-GP by Punch/Zongker.
• Csim by Shuqing Zeng.
• Csim is a collision simulator, using a
  simulated laser range finder as the sensory
  input.
• The robot is controlled by two values
• Desired heading direction.
• Desired speed.
• These are automatically mapped to the low-level
  control.
• Csim allows environment editing.

5

• Two goals
• Use GP to evolve collision avoidance ability.
• Use GP to (eventually) evolve behaviors on a
  single, non-simulated agent, in real-time.

6
Attention

• The laser scanner might return information about
  far-away objects.
• Only local information is needed to avoid nearby
  objects.
• A custom attention window removes input that is
  irrelevant to the output, by thresholding.
• I used an ellipse, with the robot at the center
• Maj. axis length 8m, Min. axis length 3m.

7
Designed Input Features

• To reduce the input dimensions to lil-gp, the
  361-d input was reduced to 4 warning
  features.
• 0 perfectly safe, 1danger.
• The 361-d input vector was broken into four
  parts, using indices 80, 180, and 280.
• The normalized average of the differences between
  the actual distance and the threshold distance
  gives the warning signal for each.
• A severe warning threshold of 1.5 was also
  set. Any single point within this threshold
  caused the respective signal to go to 1.

8
The Function Set Used

• leftwarn
• leftCenterWarn
• rightCenterWarn
• rightWarn
• Terminals that return the value of that global
  variable.
• if_greaterthan
• If the first argument is greater than the second,
  evaluate the subtree given by the third argument.
  Else evaluate the forth.
• write_heading
• write_speed
• Assign an action (from 0 to 1) to the heading or
  speed global variable.
• Ephemeral random constants from 0 to 1.

9
Simulation Setup

• Using a technique called stochasic sampling.
• After each individual is evaluated, start the
  next individual from the current state
• (position and orientation of the robot).
• Why? To try to evolve a behavior on a
  non-simulated agent, it might take much longer to
  move it back to the starting location after each
  individual.
• However, this can be unfair for some individuals.

10
Fitness

• The number of fitness cases per individual was
  set to 50.
• Evaluation only began after 25 fitness cases.
• If the agent collided with an object, the
  remaining fitness cases were set to the worst
  possible value, and that individuals turn ended.
• Fitness was otherwise calculated as the sum of
  the speed output.

11
Experiment Setup

• A current issue is speed. It takes about 5
  seconds to evaluate each individual.
• Parameters
• Max pop size 200
• Crossover 90
• Tournament selection, size7
• Reproduction 10
• Tournament selection, size7
• Init-method half and half
• Init-depth 2-6
• Max size 17

12
The Environment

• When a collision occurred, the robot was supposed
  to turn around for the next individual.
• But there was a bug due to too-high speed, the
  robot might get stuck in an obstacle.
• So, after a collision, the robot was teleported
  back to the beginning (what I didnt want to do).

13
What Evolved

• What looked like five distinct subpopulations
• (1) Some went really slow.
• (2) Some went in circles and hoped for the best.
• (3) Some went straight and hoped for the best.
• (4) Others took advantage of a
  collision-detection bug in the program.
• (5) Others actually displayed local reactive
  avoidance behavior.
• Why doesnt (5) dominate?

14
Some Were Good
15
(No Transcript)
16
Results

• The average fitness of each generation did not
  increase significantly after the second, in each
  test.
• It stays at around .45 out of 1.
• However, due to time, Ive only run these for a
  maximum of 13 iterations.
• Tree bloat was observed, however.
• Possibilities
• Run it longer.
• Use a larger population size.
• Try a different fitness function.

17
Conclusions

• This method can be use to evolve locally reactive
  collision avoidance behavior in single
  individuals.
• However, it has not yet been shown to lead to an
  entire population.
• More study is needed.