VisualMimic: Visual Humanoid Loco-Manipulation via Motion Tracking and Generation

์ €์ž: Shaofeng Yin, Yanjie Ze, Hong-Xing Yu, C. Karen Liu, Jiajun Wu | ๋‚ ์งœ: 2025-11-13 | DOI: 10.48550/arXiv.2509.20322 📄 PDF


Essence

Figure 2

Fig. 2: VisualMimic consists of two training stages: 1) training a general keypoint tracker, where a teacher motion trac

VisualMimic์€ egocentric vision๊ณผ hierarchical whole-body control์„ ๊ฒฐํ•ฉํ•œ sim-to-real ํ”„๋ ˆ์ž„์›Œํฌ๋กœ, ์ธ๊ฐ„์˜ ๋™์ž‘ ๋ฐ์ดํ„ฐ๋กœ ํ•™์Šตํ•œ task-agnostic keypoint tracker์™€ task-specific visuomotor policy๋ฅผ ํ†ตํ•ด humanoid robot์˜ loco-manipulation์„ ์‹คํ˜„ํ•œ๋‹ค.

Motivation

Achievement

Figure 3

Fig. 3: Our visuomotor policies generalize across diverse space and time, shown on the box-pushing task.

How

Figure 2

Fig. 2: VisualMimic consists of two training stages: 1) training a general keypoint tracker, where a teacher motion trac

Originality

Limitation & Further Study

Evaluation

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

์ดํ‰: VisualMimic์€ teacher-student distillation์˜ ์ฐฝ์˜์  ์ด์ค‘ ์ ์šฉ๊ณผ human motion statistics ๊ธฐ๋ฐ˜ ์ œ์•ฝ์œผ๋กœ humanoid loco-manipulation์˜ ํ˜„์‹ค์  ๊ณผ์ œ๋ฅผ ํšจ๊ณผ์ ์œผ๋กœ ํ•ด๊ฒฐํ•˜๋ฉฐ, ๋‹ค์–‘ํ•œ ์ž‘์—…์—์„œ zero-shot real-world transfer๋ฅผ ์ž…์ฆํ•œ ๋งค์šฐ ์˜๋ฏธ ์žˆ๋Š” ์—ฐ๊ตฌ์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˜์ •