[Question] How to get the local observation of each agent?

[zhaoxuesi]

Required prerequisites

• I have read the documentation https://safety-gymnasium.readthedocs.io.
• I have searched the Issue Tracker and Discussions that this hasn't already been reported. (+1 or comment there if it has.)
• Consider asking first in a Discussion.

Questions

I copy a code that writes the get_obs() as follows. However the state = self.env.task.obs() return the global observation. However, the base_task.py only give the obs(), which reture the global obs. This mean the training and execution are conducted under centrialized-training-centrialized-execution framework. But, I want the centrialized-training-decentrialized-execution framwork. Therfore, is there any methods to get the local observation of each agent?
Here is the code (get_obs() ):
_def get_obs(self):
state = self.env.task.obs() # This is global obs.
obs_n = []
for a in range(self.num_agents):
agent_id_feats = np.zeros(self.num_agents, dtype=np.float32)
agent_id_feats[a] = 1.0
obs_i = np.concatenate([state, agent_id_feats])
obs_i = (obs_i - np.mean(obs_i)) / (np.std(obs_i) + 1e-8)
obs_n.append(obs_i)
return obs_n

[muchvo]

Sorry for the late reply. Local observations can be obtained by manually indexing the global observation vector for each respective agent. This requires understanding the semantic meaning of each dimension in the observation space. MuJoCo's named access API is the recommended way to handle this. See the documentation here: MuJoCo Named Access.

[zhaoxuesi]

Thanks for your reply so much. I have the inter-look the observation vectors. The local obs vector is dafultly equal to the global obs.
So every local obs constains all of states leading that you can just dissemble the vector according to the value names. It really drives me crazy.

[zhaoxuesi]

One more question, please. Why does one of my two agent allways converges faster than than other one what ever I adjust the paremeters and seeds. It is so weird.