BabyAI is a platform for training reinforcement learning agents to understand and execute natural language commands in simulated environments. It serves as a testbed for studying grounded language acquisition and sample efficiency, targeting AI researchers and practitioners in NLP and RL.

How It Works

BabyAI provides a suite of procedurally generated environments with varying complexity, designed to test an agent's ability to interpret and act upon linguistic instructions. The core approach involves training agents using a combination of reinforcement learning and imitation learning, leveraging demonstrations generated by a rule-based "BabyAI bot" for efficient learning.

Quick Start & Requirements

• Install: Clone the repository, create a conda environment (conda env create -f environment.yaml), activate it (source activate babyai), clone gym-minigrid, install it (pip install --editable .), then install BabyAI (pip install --editable .).
• Prerequisites: Python 3.6+, OpenAI Gym, NumPy, PyTorch 0.4.1+, blosc.
• Setup: Requires cloning two repositories and setting a BABYAI_STORAGE environment variable.
• Demos: Downloadable demos are available, with an MD5 checksum provided for verification.
• Docs: Instructions on how to contribute, Codebase Structure, Training, Evaluation and Reproducing Baseline Results, BabyAI 1.0+ levels and older levels.

Highlighted Details

• Focuses on sample efficiency in grounded language acquisition.
• Supports imitation learning using bot-generated demonstrations.
• Offers a compact observation format (7x7x3) and an optional RGB pixel observation format via RGBImgPartialObsWrapper.
• Baseline results and experimental setup are detailed in associated papers (ICLR19 and arXiv:2007.12770).

Maintenance & Community

The master branch is updated frequently, but the repository is noted as not actively maintained. The BabyAI 1.1 branch is recommended for replicating baseline results.

Licensing & Compatibility

The repository does not explicitly state a license. The associated gym-minigrid repository is MIT licensed.

Limitations & Caveats

This repository is not actively maintained, with all BabyAI environments now part of the Minigrid library. Using the pixel observation architecture is not compatible with imitation learning due to how demonstrations were generated.