DoReMi: Grounding Language Model by Detecting and Recovering from Plan-Execution Misalignment

์ €์ž: Yanjiang Guo, Yen-Jen Wang, Lihan Zha, Jianyu Chen | ๋‚ ์งœ: 2023-07-01 | URL: https://arxiv.org/abs/2307.00329 📄 PDF


Essence

Figure 1

Fig. 1: Illustration of our motivation. Previous works use LLM to generate only high-level textual plans. Therefore, Low

DoReMi๋Š” LLM์œผ๋กœ ๊ณ ์ˆ˜์ค€ ๊ณ„ํš๊ณผ ์‹คํ–‰ ์ œ์•ฝ์กฐ๊ฑด์„ ๋™์‹œ์— ์ƒ์„ฑํ•˜๊ณ , VLM์œผ๋กœ ์‹คํ–‰ ์ค‘ ์ œ์•ฝ ์œ„๋ฐ˜์„ ์ง€์†์ ์œผ๋กœ ๊ฐ์ง€ํ•˜์—ฌ ๊ณ„ํš-์‹คํ–‰ ๋ถˆ์ผ์น˜๋ฅผ ์ฆ‰์‹œ ํƒ์ง€ํ•˜๊ณ  ๋ณต๊ตฌํ•˜๋Š” ๋กœ๋ด‡ ์ž‘์—… ํ”„๋ ˆ์ž„์›Œํฌ์ด๋‹ค.

Motivation

Achievement

Figure 2

Fig. 2: Previous methods perform open-loop planning or only re-plan when the previous skill is finished. Our DoReMi

How

Figure 3

Fig. 3: Open-ended scene descriptions of VLMs are ambiguous. DoReMi leverages the LLM to reason specific constraints

Originality

Limitation & Further Study

Evaluation

Novelty: 4/5 Technical Soundness: 3/5 Significance: 4/5 Clarity: 4/5 Overall: 4/5

์ดํ‰: DoReMi๋Š” LLM๊ณผ VLM์„ ์ฐฝ์˜์ ์œผ๋กœ ๊ฒฐํ•ฉํ•˜์—ฌ ๋กœ๋ด‡ ์ž‘์—…์˜ ๊ณ„ํš-์‹คํ–‰ ๋ถˆ์ผ์น˜ ๋ฌธ์ œ๋ฅผ ์ฆ‰์‹œ ๊ฐ์ง€ํ•˜๊ณ  ๋ณต๊ตฌํ•˜๋Š” ์‹ค์šฉ์ ์ธ ํ”„๋ ˆ์ž„์›Œํฌ๋ฅผ ์ œ์‹œํ–ˆ์œผ๋ฉฐ, ๋ช…ํ™•ํ•œ ๋™๊ธฐ, ์ฒด๊ณ„์ ์ธ ๋ฐฉ๋ฒ•๋ก , ๊ฒฌ์‹คํ•œ ์‹คํ—˜์„ ํ†ตํ•ด ๋†’์€ ํ•™์ˆ ์  ๊ฐ€์น˜์™€ ๋กœ๋ด‡ ์ œ์–ด ๋ถ„์•ผ์˜ ์‹ค์งˆ์  ๊ธฐ์—ฌ๋ฅผ ์ž…์ฆํ–ˆ๋‹ค.

← ๋ชฉ๋ก์œผ๋กœ ๋Œ์•„๊ฐ€๊ธฐ

๐ŸŽง Audio Overview

์ด ๋…ผ๋ฌธ ๋ฆฌ๋ทฐ๋ฅผ ํŒŸ์บ์ŠคํŠธํ˜• ์˜ค๋””์˜ค๋กœ ์ƒ์„ฑํ•ฉ๋‹ˆ๋‹ค. (Gemini ยท ํ‚ค๋Š” ๋ธŒ๋ผ์šฐ์ €์—๋งŒ ์ €์žฅ ยท ์™„์„ฑ๋ณธ์€ ์ด๋ฉ”์ผ๋กœ๋„ ์ „์†ก)
โ–ธ ๊ณ ๊ธ‰: ๊ตฌ์„ฑ ๋ฐฉํ–ฅ(๋Œ€๋ณธ ์ž‘์„ฑ ์ง€์นจ) ์ง์ ‘ ์ˆ˜์ •