# BabyAI: A Platform to Study the Sample Efficiency of Grounded Language Learning

**Authors:** Maxime Chevalier-Boisvert, Dzmitry Bahdanau, Salem Lahlou, Lucas Willems, Chitwan Saharia, Thien Huu Nguyen, Yoshua Bengio  
**arXiv:** [1810.08272](https://arxiv.org/abs/1810.08272)  
**Date:** 2018-10-18  
**Summary page:** https://www.summarizepaper.com/en/paper/1810.08272v4/

## Key points

- Study focuses on training artificial agents to understand language instructions through interactive human involvement
- Current learning methods are inefficient in terms of data usage and require substantial research efforts to improve
- Introduction of the BabyAI research platform, which includes humans in grounded language learning
- Platform consists of 19 levels of increasing difficulty, guiding the agent towards acquiring a synthetic language with rich combinatorial properties
- Heuristic expert agent provided to simulate a human teacher
- Baseline results reported and estimation made regarding amount of human involvement needed to train a neural network-based agent on BabyAI levels
- Findings suggest current deep learning methods lack sufficient sample efficiency for learning compositional languages
- Study contributes valuable insights into improving sample efficiency of grounded language learning
- Highlights need for further research in this area

## Summary

The study focuses on the challenge of training artificial agents to understand language instructions through interactive human involvement. The current learning methods are inefficient in terms of data usage, requiring substantial research efforts to improve. To address this issue, the researchers introduce the BabyAI research platform, which aims to include humans in grounded language learning. The platform consists of 19 levels of increasing difficulty, guiding the agent towards acquiring a synthetic language that is a subset of English with rich combinatorial properties. Additionally, the platform provides a heuristic expert agent to simulate a human teacher. Baseline results are reported and an estimation is made regarding the amount of human involvement needed to train a neural network-based agent on some BabyAI levels. The findings suggest that current deep learning methods lack sufficient sample efficiency when it comes to learning compositional languages. This study contributes valuable insights into improving the sample efficiency of grounded language learning and highlights the need for further research in this area.