Latest AI news 2023/12/22 ์ตœ์‹  AI ๋‰ด์Šค


Neural feels with neural fields: Visuo-tactile perception for in-hand manipulation์‹ ๊ฒฝ ํ•„๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์‹ ๊ฒฝ ๊ฐ๊ฐ: ์† ์•ˆ์—์„œ์˜ ์กฐ์ž‘์„ ์œ„ํ•œ ์‹œ๊ฐ-์ด‰๊ฐ ์ธ์‹

CMU

FAIR

UC Berkeley

TU Dresden

CeTI


Abstract

To achieve human-level dexterity, robots must infer spatial awareness from multimodal sensing to reason over contact interactions. During in-hand manipulation of novel objects, such spatial awareness involves estimating the object's pose and shape. The status quo for in-hand perception primarily employs vision, and restricts to tracking a priori known objects. Moreover, visual occlusion of objects in-hand is imminent during manipulation, preventing current systems to push beyond tasks without occlusion. We combine vision and touch sensing on a multi-fingered hand to estimate an object's pose and shape during in-hand manipulation. Our method, NeuralFeels, encodes object geometry by learning a neural field online and jointly tracks it by optimizing a pose graph problem. We study multimodal in-hand perception in simulation and the real-world, interacting with different objects via a proprioception-driven policy. Our experiments show final reconstruction F-scores of 81% and average pose drifts of 4.7,mm, further reduced to 2.3,mm with known CAD models. Additionally, we observe that under heavy visual occlusion we can achieve up to 94% improvements in tracking compared to vision-only methods. Our results demonstrate that touch, at the very least, refines and, at the very best, disambiguates visual estimates during in-hand manipulation. We release our evaluation dataset of 70 experiments, FeelSight, as a step towards benchmarking in this domain. Our neural representation driven by multimodal sensing can serve as a perception backbone towards advancing robot dexterity. Videos can be found on our project website https://suddhu.github.io/neural-feels/


์ธ๊ฐ„ ์ˆ˜์ค€์˜ ๋ฏผ์ฒฉ์„ฑ์„ ๋‹ฌ์„ฑํ•˜๊ธฐ ์œ„ํ•ด, ๋กœ๋ด‡์€ ๋‹ค์ค‘๋ชจ๋‹ฌ ๊ฐ์ง€์—์„œ ๊ณต๊ฐ„ ์ธ์‹์„ ์ถ”๋ก ํ•˜์—ฌ ์ ‘์ด‰ ์ƒํ˜ธ์ž‘์šฉ์„ ์ดํ•ดํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค. ์ƒˆ๋กœ์šด ๋ฌผ์ฒด๋ฅผ ์† ์•ˆ์—์„œ ์กฐ์ž‘ํ•˜๋Š” ๋™์•ˆ, ์ด๋Ÿฌํ•œ ๊ณต๊ฐ„ ์ธ์‹์€ ๋ฌผ์ฒด์˜ ์ž์„ธ์™€ ํ˜•ํƒœ๋ฅผ ์ถ”์ •ํ•˜๋Š” ๊ฒƒ์„ ํฌํ•จํ•ฉ๋‹ˆ๋‹ค. ์† ์•ˆ์—์„œ์˜ ์ธ์‹์— ๋Œ€ํ•œ ํ˜„ ์ƒํƒœ๋Š” ์ฃผ๋กœ ์‹œ๊ฐ์„ ํ™œ์šฉํ•˜๋ฉฐ, ์‚ฌ์ „์— ์•Œ๋ ค์ง„ ๋ฌผ์ฒด๋ฅผ ์ถ”์ ํ•˜๋Š” ๊ฒƒ์— ์ œํ•œ๋ฉ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์กฐ์ž‘ ์ค‘ ์† ์•ˆ์˜ ๋ฌผ์ฒด์— ๋Œ€ํ•œ ์‹œ๊ฐ์  ๊ฐ€๋ ค์ง์€ ๋ถˆ๊ฐ€ํ”ผํ•˜์—ฌ, ํ˜„์žฌ ์‹œ์Šคํ…œ์ด ๊ฐ€๋ ค์ง์ด ์—†๋Š” ์ž‘์—…์„ ๋„˜์–ด์„œ๋Š” ๊ฒƒ์„ ๋ฐฉํ•ดํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์† ์•ˆ์—์„œ ๋ฌผ์ฒด์˜ ์ž์„ธ์™€ ํ˜•ํƒœ๋ฅผ ์ถ”์ •ํ•˜๊ธฐ ์œ„ํ•ด ๋‹ค์ค‘ ์†๊ฐ€๋ฝ์ด ๋‹ฌ๋ฆฐ ์†์— ์‹œ๊ฐ๊ณผ ์ด‰๊ฐ ์„ผ์‹ฑ์„ ๊ฒฐํ•ฉํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๋ฐฉ๋ฒ•, NeuralFeels๋Š” ์˜จ๋ผ์ธ์—์„œ ์‹ ๊ฒฝ ํ•„๋“œ๋ฅผ ํ•™์Šตํ•˜์—ฌ ๋ฌผ์ฒด ๊ธฐํ•˜ํ•™์„ ์ธ์ฝ”๋”ฉํ•˜๊ณ , ์ž์„ธ ๊ทธ๋ž˜ํ”„ ๋ฌธ์ œ๋ฅผ ์ตœ์ ํ™”ํ•จ์œผ๋กœ์จ ์ด๋ฅผ ๊ณต๋™์œผ๋กœ ์ถ”์ ํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜๊ณผ ์‹ค์ œ ์„ธ๊ณ„์—์„œ ๋‹ค์–‘ํ•œ ๋ฌผ์ฒด์™€์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ํ†ตํ•ด ๋‹ค์ค‘๋ชจ๋‹ฌ ์† ์•ˆ ์ธ์‹์„ ์—ฐ๊ตฌํ•ฉ๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ์‹คํ—˜์€ ์ตœ์ข… ์žฌ๊ตฌ์„ฑ F-์ ์ˆ˜ 81%์™€ ํ‰๊ท  ์ž์„ธ ํŽธ์ฐจ 4.7mm๋ฅผ ๋ณด์—ฌ์ฃผ๋ฉฐ, ์•Œ๋ ค์ง„ CAD ๋ชจ๋ธ๋กœ 2.3mm๊นŒ์ง€ ์ค„์–ด๋“ญ๋‹ˆ๋‹ค. ๋˜ํ•œ, ์‹ฌ๊ฐํ•œ ์‹œ๊ฐ์  ๊ฐ€๋ ค์ง ํ•˜์—์„œ ์šฐ๋ฆฌ๋Š” ์‹œ๊ฐ-๋‹จ๋… ๋ฐฉ๋ฒ•์— ๋น„ํ•ด ์ถ”์ ์—์„œ ์ตœ๋Œ€ 94% ๊ฐœ์„ ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ๋‹ค๋Š” ๊ฒƒ์„ ๊ด€์ฐฐํ–ˆ์Šต๋‹ˆ๋‹ค. ์šฐ๋ฆฌ์˜ ๊ฒฐ๊ณผ๋Š” ์ด‰๊ฐ์ด ์ตœ์†Œํ•œ ์‹œ๊ฐ ์ถ”์ •์„ ์ •์ œํ•˜๊ณ , ์ตœ๊ณ ์˜ ๊ฒฝ์šฐ์—๋Š” ์‹œ๊ฐ ์ถ”์ •์„ ๋ช…ํ™•ํžˆ ํ•˜๋Š” ๊ฒƒ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. ์šฐ๋ฆฌ๋Š” ์ด ๋ถ„์•ผ์—์„œ ๋ฒค์น˜๋งˆํ‚น์„ ํ–ฅํ•œ ๋‹จ๊ณ„๋กœ์„œ 70๊ฐœ์˜ ์‹คํ—˜์œผ๋กœ ๊ตฌ์„ฑ๋œ ํ‰๊ฐ€ ๋ฐ์ดํ„ฐ์…‹, FeelSight๋ฅผ ๊ณต๊ฐœํ•ฉ๋‹ˆ๋‹ค. ๋‹ค์ค‘๋ชจ๋‹ฌ ๊ฐ์ง€์— ์˜ํ•ด ๊ตฌ๋™๋˜๋Š” ์šฐ๋ฆฌ์˜ ์‹ ๊ฒฝ ํ‘œํ˜„์€ ๋กœ๋ด‡ ๋ฏผ์ฒฉ์„ฑ์„ ๋ฐœ์ „์‹œํ‚ค๋Š” ๋ฐ ์žˆ์–ด ์ธ์‹์˜ ๊ธฐ๋ฐ˜์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.





3์ค„ ์š”์•ฝ


1. "NeuralFeels"๋Š” ๋‹ค์ค‘ ์†๊ฐ€๋ฝ์ด ๋‹ฌ๋ฆฐ ๋กœ๋ด‡ ์†์— ์‹œ๊ฐ๊ณผ ์ด‰๊ฐ ์„ผ์‹ฑ์„ ๊ฒฐํ•ฉํ•˜์—ฌ ์† ์•ˆ์—์„œ ๋ฌผ์ฒด์˜ ์ž์„ธ์™€ ํ˜•ํƒœ๋ฅผ ์ถ”์ •ํ•˜๋Š” ๋ฐฉ๋ฒ•์ž…๋‹ˆ๋‹ค.ย 

2. ์ด ์‹œ์Šคํ…œ์€ ์‹œ๊ฐ์  ๊ฐ€๋ ค์ง ์ƒํ™ฉ์—์„œ๋„ ํšจ๊ณผ์ ์ด๋ฉฐ, ์ด‰๊ฐ์€ ์‹œ๊ฐ ์ถ”์ •์„ ์ •์ œํ•˜๊ฑฐ๋‚˜ ๋ช…ํ™•ํžˆ ํ•  ์ˆ˜ ์žˆ์Œ์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค.ย 

3. "NeuralFeels"๋Š” ๋‹ค์–‘ํ•œ ๋ฌผ์ฒด์™€์˜ ์ƒํ˜ธ์ž‘์šฉ์„ ํ†ตํ•œ 70๊ฐœ์˜ ์‹คํ—˜ ๋ฐ์ดํ„ฐ์…‹ "FeelSight"๋ฅผ ์ œ๊ณตํ•˜๋ฉฐ, ๋กœ๋ด‡ ๋ฏผ์ฒฉ์„ฑ ํ–ฅ์ƒ์— ๊ธฐ์—ฌํ•  ์ˆ˜ ์žˆ๋Š” ๊ธฐ์ˆ ์ž…๋‹ˆ๋‹ค.




https://arxiv.org/pdf/2312.13469.pdf

https://suddhu.github.io/neural-feels/