🐛 No training results for dual drone env

[bzx20]

❓ Issue

When I train the model, I receive no response using GPU. So I set checkpoints to run the code, only to find that it exits after an execution of function reset_env. When using CPU, I will wait for a long time before it begins computing. But I still cannot get the code running to end.

I use pid policy to check my environment, and the output seems normal and records the animation like this:

The intermediate variable output for solving the equation is also normal.

🤔 Possible solutions

1. Use logging to find which step goes wrong. (I have already used jax.debug.print(), but I received no logging results while training by GPU.)

2. Check input and output of step function to see if there exits any invalid values.

[bzx20]

Test: Use only taut states, and apply thrust to balance gravity, set tau as zero.

Output

taut_action: [array(0.56196034, dtype=float32), array(0., dtype=float32), array(0.56196034, dtype=float32), array(0., dtype=float32)], no problem

Animation

It should hover, but

TODO

Check taut_dual.py

[Chaoyi Pan]

Current issue

State Definition Optimizations: The current practice uses the positions of two quadrotors to define the state. This method, however, results in the introduction of unnecessary and redundant parameters. We propose to refine this approach by using the object's position instead of the quadrotor's position.

Misinterpretation of Angles: The angles $\theta, \phi$ are currently not represented correctly. These should be characterized in the world frame for a more accurate depiction.

[Chaoyi Pan]

Test results

[Chaoyi Pan]

Training results