[RL Unplugged] - Trained policies for finger_turn_hard do not match the datasets

[Jen Alchimowicz]

Hi, thanks for the good work.

I am running custom offline evaluation on the tasks from DeepMind Control Suite Dataset. I am taking the data from Tensorflow Datasets here and downloading the maching policies from the GCP bucket here.

Most tasks work fine, but when I get to finger_turn_hard I get the following error while inferring the policies on the data:

It looks like the policy is expecting one of the keys to be "touch", however, this key is not in the observations. According to the documentation here, there is no dimension "touch" in this dataset.

For context, I am loading the policy using:

tf.saved_model.load(policy_path)

And inferring using:

if hasattr(policy, 'initial_state'):
    action = policy(observation, ((),))[0]
else:
    action = policy(observation)

Am I missing something here? How can I infer the finger_turn_hard policies on the provided datasets?