Original source code

Original: flappy-automation-test, in ROS-kinetic and python2

My demo

demo.mp4

My score

• "Good performance with strong robustness!"

• In 20 consecutive tests, the scores are as follows:

Test No.	1	2	3	4	5	6	7	8	9	10
Score	114	118	116	116	109	111	110	114	117	112

Test No.	11	12	13	14	15	16	17	18	19	20
Score	116	112	116	115	117	113	116	118	112	115

My solution

1. Position integral --> get current position

   • Exact Euler method works better than forward Euler method

2. Y-axis control --> move the bird to self.goal_y

   • At first I use bang-bang control
     • Problem: oscillation
   • Then I add PD control when it's close to self.goal_y
     • Problem: hard to estimate time cost = self.goal_y_reach_count * DELTA_T
   • Because the game is discretized in 30 FPS, I finally choose MPC for motion control
     • Benefit: faster response, more accurate time cost

3. Obstacle update --> find the gate to pass

   • Set self.pos_up and self.pos_low for upper and lower bound of the gate
   • Apply two 1-D-arrays (0/1:free/stone) to represent forward and backward obstacles it can scan
   • Data processing:
     • Raw processing --> Clustering:
       • only forward; backward + forward; only backward

4. Bird status

   • The bird has 4 status: init, free, constrained, forecast
   • Set current status
   • Take different actions for different status:
     • init: y-axis: None; x-axis: Speed up
     • free: y-axis: Move to the goal; x-axis: Adjust Vx
     • constrained: y-axis: None; x-axis: slow down to VX_SAFE
     • forecast: y-axis: None; x-axis: adjust vx to Vx_desire

5. Find the gate self.goal_y

   • The gate width should be 0.5m
   • Possible gate: detected free space width is in 0.5m < d < 0.5m(1+10%)
     • if only one gate --> pass
     • if multiple gates --> go to the closest one
     • if no such gates --> go to maximum free(unobserved) space

6. Publish Accelerations and save delta-x,y