Try to fix normal runtime error in the training phase

[MiaoDX]

Hi,

As discussed in #24

It seems G1 velocity demo is not working properly, please have a check.

As zhenwang-civil suggested, change bad_orientation = DoneTerm(func=mdp.bad_orientation, params={"limit_angle": 0.8}) from 0.8 to 0.7 does fix this problem.

But I not so sure if this the best solution, after dig into the doc and help from codex/claude, it seems we can also change the loss function def.

But still, I am new to this area, please give some commits @Agnel-Wang

Sincerely,
MiaoDX

[Agnel-Wang]

Ok, I’ll update a more stable version later. But for the current flat-ground env, the training is already completed before the program exits, so you can just pick a converged iteration directly.

[filippottaviani]

Hi, changing the limit_angle doesn't fix the issue in my case. It seams to occur randomly after 50k iterations and I can't figure out why.