ManiFlow: A General Robot Manipulation Policy via Consistency Flow Training

์ €์ž: Ge Yan, Jiyue Zhu, Yuquan Deng, Shiqi Yang, Ri-Zhao Qiu, Xuxin Cheng, Marius Memmel, Ranjay Krishna, Ankit Goyal, Xiaolong Wang, Dieter Fox | ๋‚ ์งœ: 2025-09-01 | URL: https://arxiv.org/abs/2509.01819 📄 PDF


Essence

Figure 2

Figure 2: Policy Architecture of ManiFlow. Our system processes 2D or 3D visual observations,

ManiFlow๋Š” flow matching๊ณผ consistency training์„ ๊ฒฐํ•ฉํ•˜์—ฌ 1-2 inference step์œผ๋กœ ๊ณ ํ’ˆ์งˆ์˜ dexterous action์„ ์ƒ์„ฑํ•˜๋Š” visuomotor imitation learning policy์ด๋‹ค. DiT-X ์•„ํ‚คํ…์ฒ˜๋ฅผ ํ†ตํ•ด visual, language, proprioceptive ์ž…๋ ฅ์„ ํšจ์œจ์ ์œผ๋กœ ์กฐ๊ฑดํ™”ํ•˜๋ฉฐ ์‹ค์ œ ๋กœ๋ด‡ ํ™˜๊ฒฝ์—์„œ ์šฐ์ˆ˜ํ•œ ์„ฑ๋Šฅ์„ ๋ณด์ธ๋‹ค.

Motivation

Achievement

Figure 1

Figure 1: We introduce ManiFlow, a flow matching model excelling in complex manipulation tasks,

How

Figure 3

Figure 3: ManiFlow Consistency Training. Given a flow path that smoothly transforms action

Originality

Limitation & Further Study

Evaluation

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

์ดํ‰: ManiFlow๋Š” flow matching๊ณผ consistency training์˜ ํšจ๊ณผ์ ์ธ ๊ฒฐํ•ฉ, ์ฒด๊ณ„์ ์ธ ablation ๋ถ„์„, ๊ทธ๋ฆฌ๊ณ  ํฌ๊ด„์ ์ธ ์‹ค์ œ ํ™˜๊ฒฝ ๊ฒ€์ฆ์„ ํ†ตํ•ด robot manipulation ๋ถ„์•ผ์—์„œ ์ƒ๋‹นํ•œ ์ง„์ „์„ ์ด๋ฃจ์—ˆ๋‹ค. ํŠนํžˆ inference ํšจ์œจ์„ฑ๊ณผ ์‹ค์ œ ์„ฑ๋Šฅ์˜ ๋™์‹œ ํ–ฅ์ƒ์€ ์‹ค๋ฌด ์ ์šฉ ๊ฐ€๋Šฅ์„ฑ์„ ๋†’์ด๋Š” ์ค‘์š”ํ•œ ๊ธฐ์—ฌ์ด๋‹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˜์ •