From Grounding to Manipulation: Case Studies of Foundation Model Integration in Embodied Robotic Systems

TL;DR

This paper compares different foundation model integration strategies in embodied robotics, evaluating their effectiveness in instruction understanding and manipulation tasks, and discusses design implications and future challenges.

Contribution

It provides a systematic evaluation of three FM integration paradigms in robotics through case studies, highlighting trade-offs and guiding design choices for language-driven agents.

Findings

End-to-end VLA models excel in complex instruction understanding.

Modular pipelines offer better data efficiency in skill transfer.

Trade-offs exist between generalization and data efficiency in FM-based robotics.

Abstract

Foundation models (FMs) are increasingly used to bridge language and action in embodied agents, yet the operational characteristics of different FM integration strategies remain under-explored -- particularly for complex instruction following and versatile action generation in changing environments. This paper examines three paradigms for building robotic systems: end-to-end vision-language-action (VLA) models that implicitly integrate perception and planning, and modular pipelines incorporating either vision-language models (VLMs) or multimodal large language models (LLMs). We evaluate these paradigms through two focused case studies: a complex instruction grounding task assessing fine-grained instruction understanding and cross-modal disambiguation, and an object manipulation task targeting skill transfer via VLA finetuning. Our experiments in zero-shot and few-shot settings reveal trade-offs in generalization and data efficiency. By exploring performance limits, we distill design implications for developing language-driven physical agents and outline emerging challenges and opportunities for FM-powered robotics in real-world conditions.