mpschrader / gym-sokoban Goto Github PK

Python can't import the pip installation

ModuleNotFoundError: No module named 'gym_sokoban'

maybe a typo?

Hi, this repository is currently listed in the gymnasium third party environments but we are cleaning the list up to only include maintained gymnasium-compatible repositories.

Would it be possible for it to be upgraded from gym to gymnasium? Gymnasium is the maintained version of openai gym and is compatible with current RL training libraries (rllib and tianshou have already migrated, and stable-baselines3 will soon).

For information about upgrading and compatibility, see migration guide and gym compatibility. The main difference is the API has switched to returning truncated and terminated, rather than done, in order to give more information and mitigate edge case issues. The documentation explains how to easily convert your code.

Recently DeepMind published 1.000.000 pre-generated levels (Repo). We would love to integrate these levels as a new variation.

Hey,
Can I assume there would always be a frame to every board? namely, that the right/left-most and upper/lower locations are walls?

It would be great to refactor the TinyWorld environments to not be their own classes, but to be new rendering modes.

This means two new rendering modes should be introduced:

• tiny_rgb_array
• tiny_human

Hi, in the Sokoban code, I tried running the code for max steps 150, but I see that the env files have max_steps default to 120 and the done is set to True even when target is not reached.

Is there a way I can define the max_steps to higher value in env? Is there any function or value to be set?

Hi!
I would like to know if there is any way of modifying the max time steps so that the enviorment doesnt disappear early.

Possible variations:

• Two Player
• Push and Pull
• Box Specific Targets

As for many gym environments, action 0 should map to NOOP (no operation = no movement for a step of the environment).
Some common environment wrapper and agents utilize this option, especially the paper deepmind released the boxoban-levels for:
https://deepmind.com/research/publications/investigation-model-free-planning/

Context:

1. Install gym-sokoban from the source:
   python install -e .

2. Running the following code in a Jupyter notebook:

import gym 
import gym_sokoban

env = gym.make('Sokoban-v0')
#env.reset()

env.render(mode='human')

action = env.action_space.sample()
observation, reward, done, info = env.step(action)

3. Error:

---------------------------------------------------------------------------
FileNotFoundError                         Traceback (most recent call last)
Cell In[9], line 1
----> 1 env = gym.make('Sokoban-v0')
      2 #env.reset()
      4 env.render(mode='human')

File ~/miniconda3/envs/llama3/lib/python3.9/site-packages/gym/envs/registration.py:640, in make(id, max_episode_steps, autoreset, apply_api_compatibility, disable_env_checker, **kwargs)
    637     render_mode = None
    639 try:
--> 640     env = env_creator(**_kwargs)
    641 except TypeError as e:
    642     if (
    643         str(e).find("got an unexpected keyword argument 'render_mode'") >= 0
    644         and apply_human_rendering
    645     ):

File ~/miniconda3/envs/llama3/lib/python3.9/site-packages/gym_sokoban/envs/sokoban_env_variations.py:14, in __init__(self)

File ~/miniconda3/envs/llama3/lib/python3.9/site-packages/gym_sokoban/envs/sokoban_env.py:48, in SokobanEnv.__init__(self, dim_room, max_steps, num_boxes, num_gen_steps, reset)
     44 self.observation_space = Box(low=0, high=255, shape=(screen_height, screen_width, 3), dtype=np.uint8)
     46 if reset:
     47     # Initialize Room
---> 48     _ = self.reset()
...
    408 else:
    409     # Writing: check that the directory to write to does exist
    410     dn = os.path.dirname(fn)

FileNotFoundError: No such file: '/home/xy/miniconda3/envs/llama3/lib/python3.9/site-packages/gym_sokoban/envs/surface/box.png'
Output is truncated. View as a scrollable element or open in a text editor. Adjust cell output settings...

4. After checking the site-package folder, I found there is no gym_sokoban folder. There is only an egg-link file:

gym-sokoban.egg-link

I am not sure how I can fix the above issue and get it running.

If I try to run a script twice which imports "gym_sokoban", I get the error "cannot re-register id: Sokoban-v0"

I have to restart my kernel to be able to run it again.

Any suggestions?

Possible Instructions:
https://medium.com/@joel.barmettler/how-to-upload-your-python-package-to-pypi-65edc5fe9c56

Hello! Firstly thanks for creating this repo and sokoban package!

I'm running into some initial issues in trying to sanity check the install and environment. When I try to run Random_Sampling.py, I run into an error:
"ImportError: cannot import name 'rendering' from 'gym.envs.classic_control' " , and it errors on line: env.render(mode='human').

Any thoughts on what might be causing the issue or how I could fix it?

Thanks!

how can we make the package load and environment start up quicker?

there are some nested for loops, could we potentially replace or parallelize them?

The readme file for the push and pull variant of the game has the Fixed Target Room IDs instead of the push and pull room ids:
https://github.com/mpSchrader/gym-sokoban/blob/master/docs/variations/PushAndPull.md

i read your code and i found that reward = -0.1 + sum_component_rewards. So instead of return 1 for Push Box on Target, env returns 0.9 and instead of return 10 for winning, env returns -0.1+1+10= 10.9

Sometime this warning "[SOKOBAN] Runtime Error/Warning: Generated Model with score == 0
[SOKOBAN] Retry . . ." happens. This effects the speed of game.

scipy.misc.read has been removed in latest scipy
You can just refactor to imageio, as mentioned in the reference.

Hi,

Thank you for open source this awesome project!

Is there a mechanism for avoiding runtime issue like:

env = gym.make('TinyWorld-Sokoban-small-v0')  # Error
# RuntimeError: Not enough free spots (#3) to place 1 player and 2 boxes.

There is also possible issue when we reset the environment

env = gym.make('TinyWorld-Sokoban-small-v0')
for _ in range(int(1e3)):
    env.reset()
# RuntimeError: Not enough free spots (#3) to place 1 player and 2 boxes.
# or
# RuntimeWarning: Generated Model with score == 0 

A simple workaround for the issue with reset could be:

def callback_sokoban_reset(f):
    def callback():
        try:
            return f()
        except (RuntimeWarning, RuntimeError):
            print("[SOKOBAN] Runtime error retry . . .")
            return callback()
    return callback

env = gym.make('TinyWorld-Sokoban-small-v0')
env.reset = callback_sokoban_reset(env.reset)

but there is maybe a more efficient way of dealing with these issues.

How I can get the number states of env? I try to used policy iteration and value iteration to solve this!

Hi,
could it be that in file sokoban_env.py function "_check_if_done()" is not considering self.max_steps in evaluation, if a game is done?

Create example that allows to record a game.

Hello,

Is there any RL baseline configuration for Sokoban gym environment. I want to compare to working learning agent on this game in order to see if my approach is doing good enough.

Thank you.

Hi,
function _calc_reward() in sokoban_env.py is only called, when action is < 4. So if i start training and actions are >= 4 the penalty for steps is 0 until the chosen action is < 4 for the first time. I don´t know if this has any effect in the longrun (e.g. if rewards are calculated correctly in general)

Pip version gym-Sokoban doesn't support kwargs currently. It took me 1 hour to debug looking at the newest version code on github.
TypeError: __init__() got an unexpected keyword argument 'dim_room'

Hello

I am amazed by your work. I am wondering if you tested the Sokoban's game on the standard RL method (Q learning, A2C, ec), and wondering if you have success rate for this kind of game?

I am not really familiar with the code in this Github repository, in particular, with the code that generates the rooms. However, I was told that the optimal policy for each room can be generated from the "reverse playing" algorithm that is used to ensure that the rooms are solvable. So, is this true? If yes, what's the easiest way to do it? (Of course, if I get familiar with the code, I will be able to answer this question, but if you can immediately answer it, that would save me some time: I am looking for some environments where the optimal policy is known, so I was wondering if I could use your code to compute the optimal policy).

got unexpected keyword argument 'second_player'

When using the PushAndPullSokobanEnv the init has a reset at the end, even though the base class already has a reset.
This causes env.reset() not to return to the same environment even when using constant seed.

To fix remove the extra reset on the init of PushAndPullSokobanEnv

hello!
Thanks for your code so much.I am wondering what's the difference between action 0-3 and action 4-7.I found they have the same consequence in my test.

Hi, I used "gym.make(Sokoban)" to create the environment, but there was an error that "No registered env with id: Sokoban-v0". How to deal with it?

Add argsparser to examples

since there is no special case for 0 and it is taken as a regular push it crashes - no CHANGE_COORDINATES value that fits.

To fix - change if action < 5 to elif (0 < action) and (action < 5)