The meaning of "terminated" and "truncated"

[pumfish]

Hello, I'm trying to add new simulation environment experiments in ManiSkill2-Learn.
While modifying the code, I noticed the following function in vec_env.py.

    # Special functions
    def step_dict(self, actions, idx=None, restart=True):
        idx, slice_idx = self._process_idx(idx)
        obs = self.recent_obs.slice(slice_idx).copy(wrapper=False)
        next_obs, reward, terminated, truncated, info = self.step(actions, idx=idx)
        term_or_trunc = np.logical_or(terminated, truncated)
        if np.any(term_or_trunc) and restart:
            self.reset(idx=np.where(term_or_trunc[..., 0])[0])
            
        return dict(
            obs=obs,
            next_obs=next_obs,
            actions=GDict(actions).f64_to_f32(wrapper=False),
            rewards=reward,
            dones=terminated,
            episode_dones=term_or_trunc,
            infos=info,
            worker_indices=idx,
        )

I checked the step function in ManiSkill2 and found that truncated is set to False. Doesn't that mean dones will always equal episode_dones?
Also, when a task is completed successfully and when a task reaches the maximum number of steps, what value should terminated return?

[xuanlinli17]

dones doesn't equal to episode_dones. See

ManiSkill2-Learn/tools/convert_state.py

Line 123 in e9860af

"episode_dones": False if i < length - 1 else True,

. episode_dones is merely an episode separator, while dones is success indicator. dones corresponds to terminated while episode_dones corresponds to truncated.