[Bug Report] Obs and info semantics in PointMaze with continuing_task

[younik]

Describe the bug
The observation and info returned at the last step in PointMaze with continuing_task=True, aren't updated (i.e. they contain the old goal). This is not the intended general semantics: in a common RL loop, the agent will use the old observation to predict the action to go to the old goal, instead of the new one.

See related issue: Farama-Foundation/Minari#265
See:

Gymnasium-Robotics/gymnasium_robotics/envs/maze/point_maze.py

Lines 392 to 406 in 3719d9d

	def step(self, action):
	obs, _, _, _, info = self.point_env.step(action)
	obs_dict = self._get_obs(obs)
	reward = self.compute_reward(obs_dict["achieved_goal"], self.goal, info)
	terminated = self.compute_terminated(obs_dict["achieved_goal"], self.goal, info)
	truncated = self.compute_truncated(obs_dict["achieved_goal"], self.goal, info)
	info["success"] = bool(
	np.linalg.norm(obs_dict["achieved_goal"] - self.goal) <= 0.45
	)
	# Update the goal position if necessary
	self.update_goal(obs_dict["achieved_goal"])
	return obs_dict, reward, terminated, truncated, info

Code example
You need an expert policy to see this; check https://github.com/Farama-Foundation/minari-dataset-generation-scripts/blob/main/scripts/pointmaze/create_pointmaze_dataset.py