Adding Custom Domains

This tutorial will demonstrate how to add additional domains to A2Perf.

Path to A2Perf Domains

Domains are stored in A2Perf/a2perf/domains. Currently, you’ll find the main domains in this folder such as circuit_training, quadruped_locomotion, and web_navigation. There are also additional modules such as tfa that is there for interfacing with Tensorflow Agents.

---

Creating a new domain

To add a new domain:

1. Create a new folder in the domains directory. For the purposes of this tutorial, we will create a new domain called “my_domain”:

cd A2Perf/a2perf/domains
mkdir my_domain

2. Inside of my_domain, create the following files:

• __init__.py

• [any other files needed for your domain]

3. Inside of __init__.py, you will need to register your domain with gymnasium. Here is an example of how we do this for the web navigation domain:

import gymnasium as gym
import pkg_resources

base_url_path = pkg_resources.resource_filename('a2perf.domains.web_navigation',
                                                'gwob')
base_url = f'file://{base_url_path}/'
data_dir = pkg_resources.resource_filename(
        'a2perf.domains.web_navigation.environment_generation', 'data')

gym.envs.register(
        id='WebNavigation-v0',
        entry_point=(
                'a2perf.domains.web_navigation.gwob.CoDE.environment:WebNavigationEnv'
        ),
        apply_api_compatibility=False,
        disable_env_checker=False,
        kwargs=dict(
                use_legacy_step=False,
                use_legacy_reset=False,
                data_dir=data_dir,
                base_url=base_url),
)

Note: The base_url_path, base_url, and data_dir are specific to web navigation and are not necessary for your domain.

---

Additional configurations

When adding a new domain, you may need to configure additional settings:

1. Docker configurations: In a2perf/launch/docker_utils.py, add entries for your new domain:

   a. In the get_docker_instructions function:

   docker_instructions = {
       # ... existing entries ...
       BenchmarkDomain.MY_DOMAIN.value: common_setup + [
           # Add your domain-specific Docker instructions here
           # E.g., installing dependencies, setting up the environment
       ],
   }
   

   b. In the get_entrypoint function:

   entrypoints = {
       # ... existing entries ...
       BenchmarkDomain.MY_DOMAIN.value: xm.CommandList([
           # Add your domain-specific entrypoint commands here
       ]),
   }
   
   

2. Update constants: In a2perf/constants.py, add your new domain to the BenchmarkDomain enum and update the ENV_NAMES dictionary:

   import enum
   import gin
   
   @gin.constants_from_enum
   class BenchmarkDomain(enum.Enum):
       QUADRUPED_LOCOMOTION = "QuadrupedLocomotion"
       WEB_NAVIGATION = "WebNavigation"
       CIRCUIT_TRAINING = "CircuitTraining"
       MY_DOMAIN = "MyDomain"  # Add your new domain here
   
   # ... (other existing code)
   
   ENV_NAMES = {
       # ... (existing domains)
       BenchmarkDomain.MY_DOMAIN: [
           "MyDomain-Task1-v0",
           "MyDomain-Task2-v0",
           # Add all the environment names for your new domain
       ],
   }