In stable_baselines3 I'm getting an assertion error saying my reset function cannot return a Tuple

Question

I am trying to run the chech_env function in gym (OpenAI version), however it's failing on an assertion error suggesting the environment isn't configured correctly. Specifically it's saying AssertionError: The observation returned by the reset() method should be a single value, not a tuple

I've built a custom gym environment, and currently I have an observation_space that looks like this:

from gym import spaces
# class def, etc, this is not a stand alone func
def __init__(self):
    self.camera_height = 210
    self.camera_width = 160
    # lots of environment setup

    self.action_space = spaces.Box(
            low=np.array([0, -1.0]), high=np.array([1, 1.0]),
            shape=(2,), dtype=np.float32)

    self.observation_space = spaces.Box(
            low=0, high=255,
            shape=(self.camera_height, self.camera_width, 3),
            dtype=np.uint8)

And I'm returning from a get_obs() function, which is just returning an empty array of zeros for now.

def get_ob(self):
    state = np.zeros((self.camera_width, self.camera_height, 3), np.uint8)

    return state

And then those results are return directly in the return for reset(self), as such:

def reset(self):
    # lots of set up code here

    self.obs = self.get_ob()
    return self.obs, {'arg1': 0}

And I'm running the check_env function, which should test to make sure everything is working

import gym
import my_custom_env
from stable_baselines3.common.env_checker import check_env

env = gym.make('my_custom_env-v0')

check_env(env)

However it keeps complaining that it shouldn't be a Tuple, which is confusing since there are no Tuples in my code at all. The full error looks like this:

Traceback (most recent call last):
  File "check_env.py", line 10, in <module>
    check_env(env)
  File "/home/cannon-art/.local/lib/python3.8/site-packages/stable_baselines3/common/env_checker.py", line 400, in check_env
    _check_returned_values(env, observation_space, action_space)
  File "/home/cannon-art/.local/lib/python3.8/site-packages/stable_baselines3/common/env_checker.py", line 246, in _check_returned_values
    _check_obs(obs, observation_space, "reset")
  File "/home/cannon-art/.local/lib/python3.8/site-packages/stable_baselines3/common/env_checker.py", line 162, in _check_obs
    assert not isinstance(
AssertionError: The observation returned by the `reset()` method should be a single value, not a tuple

Printing out self.observation.shape and state.shape give me the same sizes and they look the same as well. I'm really at a loss here and would appreciate any thoughts you may have. Thanks!!

score 0 · Answer 1 · answered Jun 14 '23 at 21:01

I found the issue. It turns out the check_env(env) doesn't expect an info value when it calls reset. So what was happening was it was combining my data and info into a Tuple, and then complaining about it being a Tuple.

This does seem peculiar to me since in the Gym docs it calls for both observation and info, however this check is from stable_baselines3.

Of course in the stable_baselines3 documentation it returns an info field as well, which confuses me. I guess this is just further proof that you need to pay attention to your error messages, not your docs lol

In stable_baselines3 I'm getting an assertion error saying my reset function cannot return a Tuple

1 Answers1