Latent Action Diffusion for Cross-Embodiment Manipulation

์ €์ž: Erik Bauer, Elvis Nava, Robert K. Katzschmann | ๋‚ ์งœ: 2025-06-17 | URL: https://arxiv.org/abs/2506.14608 📄 PDF


Essence

Figure 1

Fig. 1: Overview of our approach. Left: We construct a semantically aligned latent action space by training modality-spe

๋กœ๋ด‡์˜ ๋‹ค์–‘ํ•œ end-effector ๊ฐ„ action space ์ด์งˆ์„ฑ์„ ๊ทน๋ณตํ•˜๊ธฐ ์œ„ํ•ด contrastive learning์œผ๋กœ ํ•™์Šต๋œ shared latent action space์—์„œ diffusion policy๋ฅผ ํ•™์Šตํ•˜์—ฌ cross-embodiment ์กฐ์ž‘์„ ์‹คํ˜„ํ•œ๋‹ค.

Motivation

Achievement

Figure 4

Fig. 4: Success rates for three different tasks comparing single-embodiment diffusion policies to cross-embodied latent

How

Figure 2

Fig. 2: The three-stage process for learning the cross-embodiment latent action space. Stage 1: Aligned end-effector (EE

Originality

Limitation & Further Study

Evaluation

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

์ดํ‰: Cross-embodiment ๋กœ๋ด‡ ํ•™์Šต์˜ action space ์ด์งˆ์„ฑ ๋ฌธ์ œ๋ฅผ learned latent representation์œผ๋กœ ์šฐ์•„ํ•˜๊ฒŒ ํ•ด๊ฒฐํ•˜๊ณ , contrastive learning๊ณผ diffusion policy๋ฅผ ์กฐํ•ฉํ•˜์—ฌ ์‹ค์ œ ์„ฑ๋Šฅ ํ–ฅ์ƒ์„ ์ž…์ฆํ•œ ๊ฐ€์น˜์žˆ๋Š” ์—ฐ๊ตฌ์ด๋‹ค. ๋‹ค๋งŒ embodiment ๋‹ค์–‘์„ฑ ๋ฒ”์œ„ ํ™•๋Œ€์™€ alignment ๋ฉ”์ปค๋‹ˆ์ฆ˜์˜ ๋” ๊นŠ์€ ๋ถ„์„์ด ํ›„์† ๊ณผ์ œ์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˜์ •