Streaming Flow Policy: Simplifying diffusion/flow-matching policies by treating action trajectories as flow trajectories

์ €์ž: Sunshine Jiang, Xiaolin Fang, Nicholas Roy, Tomรกs Lozano-Pรฉrez, Leslie Pack Kaelbling, Siddharth Ancha | ๋‚ ์งœ: 2025-05-28 | URL: https://arxiv.org/abs/2505.21851 📄 PDF


Essence

Figure 1

Figure 1: (a) Diffusion policy [1] and flow-matching policy [2] input a history of observations (not shown) to

Action trajectory๋ฅผ flow trajectory๋กœ ์ทจ๊ธ‰ํ•˜์—ฌ diffusion/flow-matching ์ •์ฑ…์„ ๋‹จ์ˆœํ™”ํ•˜๊ณ , ํ๋ฆ„ ์ƒ˜ํ”Œ๋ง ์ค‘ ์‹ค์‹œ๊ฐ„์œผ๋กœ ๋กœ๋ด‡์— action์„ ์ŠคํŠธ๋ฆฌ๋ฐํ•  ์ˆ˜ ์žˆ๋Š” streaming flow policy๋ฅผ ์ œ์•ˆํ•œ๋‹ค.

Motivation

Achievement

Figure 2

Figure 2: (a) To illustrate our method, we consider a toy example of 1-D robot actions with two demonstration

How

Figure 2

Figure 2: (a) To illustrate our method, we consider a toy example of 1-D robot actions with two demonstration

Originality

Limitation & Further Study

Evaluation

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

์ดํ‰: ๋ณธ ๋…ผ๋ฌธ์€ action trajectory๋ฅผ flow trajectory๋กœ ์ทจ๊ธ‰ํ•˜๋Š” ๊ทผ๋ณธ์ ์œผ๋กœ ์ƒˆ๋กœ์šด ๊ด€์ ์„ ์ œ์‹œํ•˜์—ฌ diffusion/flow policy์˜ ๊ณ„์‚ฐ ํšจ์œจ์„ฑ๊ณผ ๋ฐ˜์‘์„ฑ์„ ํฌ๊ฒŒ ๊ฐœ์„ ํ•œ ๋…ผ๋ฌธ์ด๋‹ค. Streaming generation์ด๋ผ๋Š” ์‹ค์šฉ์  ์ด์ ๊ณผ ์ด๋ก ์  ๊ธฐ๋ฐ˜(flow matching)์˜ ์กฐํ™”, ๊ทธ๋ฆฌ๊ณ  ๋กœ๋ด‡ ์ œ์–ด์˜ ํŠน์„ฑ์„ ํ™œ์šฉํ•œ ์„ค๊ณ„๊ฐ€ ๋‹๋ณด์ด๋Š” ์šฐ์ˆ˜ํ•œ ์—ฐ๊ตฌ๋‹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˜์ •