Yichi_zhang Danli Deliberative Agent for Following Natural Language Instructions 2022

[TOC]

1. Title: DANLI: Deliberative Agent for Following Natural Language Instructions
2. Author: Yichi Zhang
3. Publish Year: 22 Oct, 2022
4. Review Date: Sun, Nov 20, 2022

Summary of paper

Motivation

• reactive agent simply learn and imitate behaviours encountered in the training data
• these reactive agents are insufficient for long-horizon complex tasks. To address this limitation, we propose a neuro-symbolic deliberative agent that, while following language instructions, proactively applies reasoning and planning based on its neural and symbolic representations acquired from the past experience.

Contribution

• We show that our deliberative agent achieves greater than 70% improvement over reactive baselines on the challenging TEACh benchmark

Some key terms

Natural language instruction following with embodied AI agents

• parent class of language reward shaping AI agent.
• where an agent must interpret human language commands to perform actions in the physical world and achieve a goal.
• especially challenging is the hierarchical nature of everyday tasks, which often require reasoning about subgoals and reconciling them with the world state and overall goal.
• however, despite recent progress, past approaches are typical reactive in their execution of actions: conditioned on the rich, multimodal inputs from the environment, they perform actions directly without using an explicit representation of the world to facilitate grounded reasoning and planning.
• Such an approach is inefficient, as natural language instructions often omit trivial steps that a human may be assumed to already know
• Besides, the lack of any explicit symbolic component makes such approaches hard to interpret, especially when the agent make errors

DANLi architecture

1. build a uniquely rich semantic spatial representation, acquired online from the surrounding environment and language descriptions to capture symbolic information about object instances and their physical states.
2. to capture the highest level of hierarchy in tasks, we propose a neural task monitor that learns to extract symbolic information about task progress and upcoming subgoals from the dialog and action history.
3. Using these elements as a planning algorithm to plan low-level actions for subgoals in the environment, taking advantage of DANLI’s transparent reasoning and planning pipeline to detect and recover from errors.

Architecture diagram

\