- Original Pdf: pdf
- Keywords: agent, language, 3D, simulation, policy, instruction, transfer
- TL;DR: Transfer learning from powerful text-based language models makes an agent more robust to human instructions in a 3D simulated world.
- Abstract: Recent work has described neural-network-based agents that are trained to execute language-like commands in simulated worlds, as a step towards an intelligent agent or robot that can be instructed by human users. However, the instructions that such agents are trained to follow are typically generated from templates (by an environment simulator), and do not reflect the varied or ambiguous expressions used by real people. We address this issue by integrating language encoders that are pretrained on large text corpora into a situated, instruction-following agent. In a procedurally-randomized first-person 3D world, we first train agents to follow synthetic instructions requiring the identification, manipulation and relative positioning of visually-realistic objects models. We then show how these abilities can transfer to a context where humans provide instructions in natural language, but only when agents are endowed with language encoding components that were pretrained on text-data. We explore techniques for integrating text-trained and environment-trained components into an agent, observing clear advantages for the fully-contextual phrase representations computed by the well-known BERT model, and additional gains by integrating a self-attention operation optimized to adapt BERT's representations for the agent's tasks and environment. These results bridge the gap between two successful strands of recent AI research: agent-centric behavior optimization and text-based representation learning.