OPEx: A Comprehensive Analysis of LLM-Centric Agents in Embodied Instruction Following
Основные понятия
LLM-centric design significantly enhances EIF performance, pinpointing visual perception and low-level action execution as crucial bottlenecks.
Аннотация
OPEx introduces a comprehensive framework delineating essential components for solving embodied learning tasks: Observer, Planner, and Executor. Extensive evaluations reveal that LLM-centric design markedly improves EIF outcomes. The study identifies visual perception and low-level action execution as critical bottlenecks. Integrating a multi-agent dialogue strategy further elevates task performance by simplifying decision-making processes. The OPEx framework aims to optimize embodied learning agents efficiently.
OPEx
Статистика
OPEx achieves 17.74% absolute gain in SR on test seen split.
LLM-based executor is responsible for implementing plans with skill library.
FILM outputs low-level navigation and interaction actions solely with a deterministic policy.
OPEx requires significantly less in-domain data compared to FILM.
Цитаты
"LLM-centric design markedly improves EIF outcomes."
"Integrating a multi-agent dialogue strategy simplifies decision-making processes."
"The OPEx framework aims to optimize embodied learning agents efficiently."