README.md

About this project

• Create a small scale version of Ada that can be run on the Mila cluster with simple environments such as BabyAI or Crafter or Procthor.
• We can use this to test out modifications to Ada.
  • In particular, we are interested in augmenting transformer architectures with memory:
    • Memorizing transformers
    • Retrieval-Augmented Reinforcement Learning
    • Discrete Key-Value Bottleneck
  • And also getting a working version of Muesli set up (and ultimately distributed):
    • Muesli

Main files in Minigrid

To run a test env just run from the Minigrid folder:

python minigrid/manual_control.py

The CustomEnv is defined in Minigrid/minigrid/envs/customv1.py We have OOP objects that work with CustomEnv defined in Minigrid/minigrid/core/world_object_custom.py Finally, the code for generating new environments can be found in Minigrid/minigrid/templates

Current Progress

Jonathan Lim wrote up a summary doc of his finding on BabyAI

Jonathan Lim is finishing up a set up for procedural generation in BabyAI/Minigrid that will allow us to define level layouts via spreadsheets and stitch them together using randomized procedural generation techniques.

We have added a custom environment to Minigrid that rewrites the core step function to allow for object oriented mechanic creation. Objects for this environment should extend the WorldObjCustom class, and have the added callbacks:

def test_overlap(self, env: CustomV1Env, obj: WorldObj) -> bool:
        """Can this overlap with the given object? Assume this is symmetric"""

def stepped_on(self, env: CustomV1Env, approach_position:Point) -> None:
        """Runs when agent is on the same tile as this object"""

def do_pickup(self, env: CustomV1Env) -> None:
        """Runs when agent picks up this object"""

def do_dropped(self, env: CustomV1Env) -> None:
        """Runs when agent drops this object"""

def step(self, env: CustomV1Env) -> None:
        """Runs each environment step"""

Using these new callbacks we can produce a bunch of new mechanics. For example, a PushBox can be made that is pushed by the agent, but will stop when it hits walls, etc:

How to config.yaml

How to provide paths to templates

Where? -> Under layout_connor section in Minigrid/minigrid/templates/config.yaml.

Your path can be a .csv file or a directory containing .csv files. You can provide one or several paths. Paths must come in pairs of path and value:

layout_connor:
  - path: /path/to/some/template.csv
    value: 1.0
  - path: /path/to/some/directory/
    value: 2.0

Higher value means a higher chance of the path being used in a map generation. Values do not have to sum up to 1.

Assigning a value to a directory means each .csv file in the directory will be assigned the same value. For example, assuming /path/to/some/directory/ has 3 files: template_1.csv, template_2.csv, template_3.csv, then this

layout_connor:
  - path: /path/to/some/directory/
    value: 2.0

and this

layout_connor:
  - path: /path/to/some/directory/template_1.csv
    value: 2.0
  - path: /path/to/some/directory/template_2.csv
    value: 2.0
  - path: /path/to/some/directory/template_3.csv
    value: 2.0

yield the same result.

Value must be a non-negative number. Assigning 0.0 to a path exludes the path:

layout_connor:
  - path: /some/path/to/a/template.csv
    value: 1.0
  - path: /some/path/to/a/directory/
    value: 2.0
  - path: /some/exluded/path
    value: 0.0

Your paths can be repeated. If a path is repeated with a new value, its value will be overwritten:

layout_connor:
  - path: /path/to/some/template.csv
    value: 1.0
  - path: /path/to/some/template.csv
    value: 2.0

now /path/to/some/template.csv has a value of 2.0.

Overwriting of values is useful, for example, in case you want to assign a different value to a file within a directory:

layout_connor:
  - path: /some/path/to/directory_1
    value: 1.0
  - path: /some/path/to/directory_1/template_1.csv
    value: 2.0

will result in all .csv files in /some/path/to/directory_1 receiving a value of 1, except for /some/path/to/directory_1/template_1.csv that will receive 2.

How to run

Map generation

A map is generated procedurally given the templates (csv files) you have provided.

1. Change paths to templates as described in the instruction. Take note of 1, 2, 3, 4
2. Run python Minigrid/minigrid/templates/template_generator.py

Manually controled run

1. Change paths to templates as described in the instruction. Take note of 1, 2, 3, 4
2. Run python Minigrid/minigrid/manual_control.py