A2Perf: Real-World Autonomous Agents Benchmark

A2Perf is a benchmark for evaluating agents on sequential decision problems that are relevant to the real world. This repository contains code for running and evaluating participant's submissions on the benchmark platform.

Environments

A2Perf provides benchmark environments in the following domains:

• Web Navigation - This environment facilitates the creation of compositional tasks represented by dependency graphs, where automatically generated websites are completed by the trained agent.
• Quadruped Locomotion - This quadruped locomotion environment aims to teach a legged robot with 18 degrees of freedom to replicate animal-like behaviors by imitating real-world motion data to develop a diverse repertoire of skills.
• Circuit Training - Chip floorplanning, a complex and traditionally manual process, has been addressed by Google's open-source Circuit Training framework, which uses reinforcement learning to optimize chip layouts for multiple objectives.

Installation

A2Perf can be installed on your local machine:

git clone https://github.com/Farama-Foundation/A2Perf.git
cd A2Perf
git submodule sync --recursive
git submodule update --init --recursive
pip install -e .[all]

Specific Package installation

To install specific packages, you can use the following commands:

pip install -e .[web_navigation]
pip install -e .[quadruped_locomotion]
pip install -e .[circuit_training] && python setup.py circuit_training

Both x86-64 and Arch64 (ARM64) architectures are supported.
Please note that the Windows version is not as well-tested as Linux and macOS versions. It can be used for development and testing but if you want to conduct serious ( time and resource-extensive) experiments on Windows, please consider using Docker or WSL with Linux version.

API

Environments in A2Perf are registered under the names WebNavigation-v0, QuadrupedLocomotion-v0, and CircuitTraining-v0. For example, you can create an instance of the WebNavigation-v0 environment as follows:

import gymnasium as gym

from a2perf.domains import web_navigation

env = gym.make("WebNavigation-v0", difficulty=1, seed=0)

User Submission

A beginners guide to benchmarking with A2Perf is described here.

• Users can pull the template repository at https://github.com/Farama-Foundation/a2perf-benchmark-submission
  • The submission repository must include:
    • train.py - defines a global train function with the following signature:

      def train():
        """Trains the user's model."""

    • inference.py - defines the following functions:

      def load_policy(env):
        """Loads a trained policy model from the specified directory."""
      def infer_once(policy, observation):
        """Runs a single inference step using the given policy and observation."""
      def preprocess_observation(observation):
        """Preprocesses a raw observation from the environment into a format compatible with the policy."""

    • requirements.txt - lists the required Python packages and their versions for running the user's code
    • __init__.py - an empty file that allows the submission to be imported as a Python module

Gin Configuration Files

Under a2perf/submission/configs, there are default gin configuration files for training and inference for each domain. These files define various settings and hyperparameters for benchmarking.

Here's an example of an inference.gin file for web navigation:

# ----------------------
# IMPORTS
# ----------------------
import a2perf.submission.submission_util
import a2perf.domains.tfa.suite_gym

# ----------------------
# SUBMISSION SETUP
# ----------------------
# Set up submission object
Submission.mode = %BenchmarkMode.INFERENCE
Submission.domain = %BenchmarkDomain.WEB_NAVIGATION
# Submission.run_offline_metrics_only = True
Submission.measure_emissions = True

####################################
# Set up domain
####################################

####################################
# Set up benchmark mode
####################################
Submission.num_inference_steps = 10000
Submission.num_inference_episodes = 100
Submission.time_participant_code = True

# ----------------------
# SYSTEM METRICS SETUP
# ----------------------
# Set up codecarbon for system metrics
track_emissions_decorator.project_name = 'a2perf_web_navigation_inference'
track_emissions_decorator.measure_power_secs = 1
track_emissions_decorator.save_to_file = True  # Save data to file
track_emissions_decorator.save_to_logger = False  # Do not save data to logger
track_emissions_decorator.gpu_ids = None  # Enter a list of specific GPU IDs to track if desired
track_emissions_decorator.log_level = 'info'  # Log level set to 'info'
track_emissions_decorator.country_iso_code = 'USA'
track_emissions_decorator.region = 'Massachusetts'

Baselines

Baselines for all tasks are provided and are described in the article supporting A2Perf.

Environment Versioning

A2Perf keeps strict versioning for reproducibility reasons. All environments end in a suffix like "-v0". When changes are made to environments that might impact learning results, the number is increased by one to prevent potential confusion. This is follows the Gymnasium conventions.

Citation

You can cite A2Perf as:
TODO

@misc{ADD CITATION,
}