How to create two indentical benchmark instances?

It seems like the MT10/ML10 benchmarks (maybe others as well) initialize the goal positions internally every time a benchmark is called. It would be nice if there was a way to create two identical benchmark instances, for example by passing a seed to the environment as an argument.

Currently, I get:

>>> import pickle
>>> import metaworld
>>> mt10_1, mt10_2 = metaworld.MT10(), metaworld.MT10()
>>> pickle.loads(mt10_1.train_tasks[0].data)['rand_vec']
array([-0.05115513,  0.63916668,  0.02      ,  0.07276333,  0.86412829,
        0.12647354])
pickle.loads(mt10_2.train_tasks[0].data)['rand_vec']
>>> pickle.loads(mt10_2.train_tasks[0].data)['rand_vec']
array([ 0.09087166,  0.69323066,  0.02      , -0.0617444 ,  0.81741752,
        0.14243739])

Would be useful if something like the following was possible:

import pickle

import numpy as np
import metaworld

seed = 10
mt10_1, mt10_2 = metaworld.MT10(seed=seed), metaworld.MT10(seed=seed)
rand_vec_1 = pickle.loads(mt10_1.train_tasks[0].data)['rand_vec']
rand_vec_2 = pickle.loads(mt10_2.train_tasks[0].data)['rand_vec']
np.testing.assert_equal(rand_vec_1, rand_vec_2)

[ryanjulian]

@hartikainen does pickling/unpicking and/or copy.deepcopy achieve this right now? It won’t get you cross-experiment uniformity but it is a step.

@avnishn is this FR tracked in an issue?

[avnishn]

Oh sorry about that I linked the wrong issue. I meant #147 my b.