Errors when running in mujoco robotics envrionment, due to the obs dimensionality.

[WuXinyang2012]

Hi, thanks for your elegant and clear implementations of SAC and TD3. Your algorithms work quite well in all experiments except the robotics environment.

When I run your algorithm in robotics env, it will return error:

  File "/home/rfa-xw/spinningup/spinup/algos/sac/sac.py", line 139, in sac
    obs_dim = env.observation_space.shape[0]
TypeError: 'NoneType' object is not subscriptable

That's because in robotics environment, the observation is a 10 dimensional dict containing 'observation', 'desired_goal' and 'achieved_goal'. So I changed following in sac.py line 140:

    #obs_dim = env.observation_space.shape[0]
    obs_dim = env.observation_space.spaces['observation'].shape[0]

And also in line 258 for the replay buffer:

        # Store experience to replay buffer
        replay_buffer.store(o['observation'], a, r, o2['observation'], d)

In line 218 for session:

        return sess.run(act_op, feed_dict={x_ph: o['observation'].reshape(1,-1)})[0]

For the full script: the modified sac.py

I hope this could help anyone who wants to play in robotics environment!

[Bhaney44]

This is really brilliant work. I ran into the same problem, but I still cannot resolve the issue. How were you able to run the modified sac.py file in the CarRacing-v0 environment? I posted my question in late December here. Would appreciate any help!

[pringithub]

I'm getting the same error for text environments .. how has this not been changed?

[Bhaney44]

@pringithub As far as I know the problem hasn't been fixed.