Metaworld
Metaworld copied to clipboard
Objects move slightly after reset
Currently the CI fails on one of the tests where the observation and goal are compared after sequential resets. After changing to MuJoco from mujoco-py it seems that there are a few environments where issues arise because objects aren't in exactly the same places after resets. The current fix is to replace len(np.unique(np.array(task_initial_pos), axis=0)) > 1
with len(np.unique(np.array(task_initial_pos), axis=0)) > 1 and not np.allclose( task_initial_pos[0], task_initial_pos[1], rtol=1e-2, atol=1e-2 )
but this masks the problem just to get the CI to pass