Learned policy cannot transfer to Pybullet-based environment

[jc-bao]

Possible reasons

• the difference in dynamics.
  • Our case: simplified control model. Consider as a point of mass.
  • Pybullet: real rigid body.
• PID controller difference.
  • Our case: assume PID controller for angle control
  • Pybullet: running actual physics. Also, call the DSL PID controller to control the drone.

[jc-bao]

Control difference comparison

• Test case: 10Hz PID controller, keep thrust to hold z position; roll set to $\pi/24$ for 1s and $-\pi/24$ for 2s followed by $\pi/24$ for 1s.

• Results: the same

  

• Test case: same case, but change pitch

• Results: the same

  

• Test case: same case, but change both

• Results: the same

  

• Conclusion: for simple move forward/backward task, two environments has basically the same performance.

[jc-bao]

Trajectory tracking test bench

• Test case: tracking a circular trajectory

[jc-bao]

Possible issues

• Yaw changes during the control (The error is too large. )
• observation error. (By comparing the difference between two trajectory, finally located problem in trajectory normalization. ✅ )