correctness and speed of TriFinger._safety_torque_check about causalworld

[luator]

Thanks a lot for tracking this down!

Regarding fix 1:
I did some quick benchmark and can confirm that np.clip is indeed much slower than calling np.minimum and np.maximum (factor 8 on my machine...), so fixing this makes a lot of sense. From a discussion on the numpy repo I learned that there is also np.core.umath.clip which has overall best performance, so I would suggest using that instead of np.minimum/maximum (it also gives more readable code).

Regarding fix 2: I'm not sure about this change, let's discuss it in the meeting tomorrow.

[michaelfeil]

Thanks for your comment! I was not aware of np.core.umath.clip. It seems like from numpy version 1.16 to 1.17 this got ~11 times slower

timeit.timeit("np.clip(a, -3, +3)",setup="import numpy as np;a=np.asarray([1,1,2,4,-5])", number=1000000)

np.core.umath.clip is not part of the public api of numpy and not available for versions before numpy 1.17. It has comparable speed to np.minimum/maximum (around 5% faster) and is more readable.

For speed across versions, I think the np.minimum/maximum might be still the way to go.

[luator]

Hm, how about something like

try:
    clip = np.core.umath.clip
except AttributeError:
    clip = np.clip

at the top of the file? For <=1.16 np.clip should still be good and for later versions there is np.core.umath.clip.

I'd really prefer to avoid minimum/maxium as I think it makes the code much less readable (and thus easier to miss mistakes).

[michaelfeil]

I think this is a good idea!
As this also happens on 3 other parts in CausalWorld (

CausalWorld/causal_world/envs/robot/observations.py

Line 260 in 7e2958a

return np.clip(observation, self._low_norm, self._high_norm)

,

CausalWorld/causal_world/envs/scene/observations.py

Line 260 in 7e2958a

return np.clip(observation, self._low_norm, self._high_norm)

and

CausalWorld/causal_world/envs/robot/action.py

Line 96 in 7e2958a

return np.clip(action, -1.0, 1.0)

), which are called for every step()

I would add your proposal to the utils.
https://github.com/rr-learning/CausalWorld/blob/7e2958a1db3af4f801d6d6df4fd1e507304ca537/causal_world/utils/env_utils.py

[michaelfeil]

Is resolved, the logic with 2x clipping works as damping for higher velocities and #102