Curieux.JY
  • JungYeon Lee
  • Post
  • ๐Ÿ•ธ๏ธ Graph
  • Lecture
  • Note

On this page

  • ๐Ÿ” Ping Review
  • ๐Ÿ”” Ring Review
    • ํ•œ ์ค„๋กœ ์‹œ์ž‘ํ•˜๋ฉด
    • ๋ฐฐ๊ฒฝ: placement๋Š” ์™œ ์ˆจ์€ ๊ธฐ๋ณธ๊ฐ’์ด ๋๋‚˜ (ยง1ยทยง2)
    • ์—ฐ๊ตฌ ์งˆ๋ฌธ (ยง3.1)
    • ๊ณผ์ œ์™€ embodiment (ยง3.2, ๋ถ€๋ก .2)
    • ๋ฐฉ๋ฒ• ์ƒ์„ธ 1: teacherโ€“student distillation (ยง3.3, ๋ถ€๋ก .3)
    • ๋ฐฉ๋ฒ• ์ƒ์„ธ 2: taxel ํ›„๋ณด์™€ placement ๊ทœ์น™ (ยง3.4, ๋ถ€๋ก .5)
    • ๋ฐฉ๋ฒ• ์ƒ์„ธ 3: tensor layout๊ณผ encoder (ยง3.5, ๋ถ€๋ก .4)
    • ์ง๊ด€: 0์€ ์ •๋ณด๊ฐ€ ์•„๋‹ˆ๋‹ค
    • ์‹คํ—˜: ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ (ยง4.1, ๋ถ€๋ก .6)
    • ์‹คํ—˜ 1: In-Palm Rotation์˜ placement (ยง4.2, Table 1)
      • ๋ถ€๋ก .7.1: ์ ‘์ด‰ coverage (Table 10)
      • ๋ถ€๋ก .7.2: ์—ํ”ผ์†Œ๋“œ ๊ฒฐ๊ณผ (Table 11)
    • ์‹คํ—˜ 2: Grasp-and-Lift์—์„œ์˜ placement (ยง4.2, Table 2)
      • ๋ถ€๋ก .7.3: ์ ‘์ด‰ ์ง‘์ค‘๋„ (Table 12)
    • โ€œuniform์ด taxel ์ˆ˜๋ฅผ ๊ณผ๋Œ€ํ‰๊ฐ€ํ•œ๋‹คโ€๋Š” ์ฃผ์žฅ์˜ ๊ทผ๊ฑฐ ์กฐ๊ฑด
    • ์‹คํ—˜ 3: ํ‘œํ˜„ (ยง4.3, Table 3)
    • ์‹คํ—˜ 4: ํ•™์Šต ๋น„์šฉ (ยง4.4, Table 4)
    • ๊ฒฐ๋ก ๊ณผ reporting checklist (ยง5, ๋ถ€๋ก .1)
    • ์„ค๊ณ„ ๋ ˆ์Šจ: ๋…ผ๋ฌธ์ด ๋งํ•œ ๊ฒƒ๊ณผ ํ‘œ๊ฐ€ ๋ฐ›์น˜๋Š” ๋ฒ”์œ„
    • ๋น„ํŒ์ ์œผ๋กœ ๋ณด๋ฉด
      • ๊ฐ•์ 
      • ์•ฝ์ ยทํ•œ๊ณ„
    • ๊ด€๋ จ ์—ฐ๊ตฌ์™€์˜ ์ž๋ฆฌ ๋งค๊น€
    • ์š”์•ฝ
    • ์ด ๊ธ€์— ๋Œ€ํ•ด

๐Ÿ“ƒWhich Taxels ๋ฆฌ๋ทฐ

tactile
dexterity
sensor-design
representation
distillation
Which Taxels, Not How Many: Tactile Sensor Placement for Dexterous Policies
Published

November 19, 2026

  • Paper Link (arXiv:2611.10647v1)

  • Taeyun Kim*, Jinsoo Shin, Taehwa Kim, Taeyoon Kim, Jinyoung Choi (*๊ต์‹ ์ €์ž)

  • NAVER LABS, arXiv preprint (cs.RO), 2026

  1. ๐Ÿ’ก ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์ด‰๊ฐ policy ์—ฐ๊ตฌ๋Š” taxel์„ ์† ํ‘œ๋ฉด์— uniformํ•˜๊ฒŒ ๊นŒ๋Š” ๊ฒƒ์„ ๊ฒ€์ฆ ์—†์ด ๊ธฐ๋ณธ๊ฐ’์œผ๋กœ ์จ ์™”๋‹ค. ์ด ๋…ผ๋ฌธ์€ teacherยทํ•™์Šต ์ ˆ์ฐจยทํ›„๋ณด taxel 2000๊ฐœ๋ฅผ ๋ชจ๋‘ ๊ณ ์ •ํ•˜๊ณ  taxel์„ ๊ณ ๋ฅด๋Š” ๊ทœ์น™๊ณผ ์ž…๋ ฅ ํ‘œํ˜„๋งŒ ๋ฐ”๊ฟ”, ๊ทธ ๊ธฐ๋ณธ๊ฐ’์ด ๋ฌด์—‡์„ ์ˆจ๊ธฐ๊ณ  ์žˆ๋Š”์ง€ ์žฐ๋‹ค.
  2. โš™๏ธ ์ด‰๊ฐ์„ ๋ณด์ง€ ์•Š๋Š” PPO teacher์˜ rollout์—์„œ taxel๋ณ„ ์ ‘์ด‰ ํ†ต๊ณ„๋ฅผ ๋ชจ์œผ๊ณ , ๊ทธ ํ†ต๊ณ„๋กœ taxel B๊ฐœ(10~300)๋ฅผ ๊ณ ๋ฅธ ๋’ค(Top-contactยทContact-coverยทLink-proportional, ๋Œ€์กฐ๊ตฐ Least-contact) uniform๊ณผ ๋น„๊ตํ•œ๋‹ค. ํ•™์ƒ policy๋Š” DAgger๋กœ ๊ฐ™์€ teacher์—์„œ distillํ•˜๋ฉฐ, ํ‘œํ˜„ ์‹คํ—˜์—์„œ๋Š” 2000๊ฐœ ์ „๋ถ€๋ฅผ ์“ด ์ฑ„ tensor layout(RandomยทBandยทGrid)๊ณผ encoder(CNNยทLSTMยทLink-wise MLP)๋งŒ ๋ฐ”๊พผ๋‹ค.
  3. ๐ŸŽฏ XHand1 In-Palm Rotation(LSTM, 3 seeds)์—์„œ ์ ‘์ด‰ ํ†ต๊ณ„๋กœ ๊ณ ๋ฅธ 25๊ฐœ(Top-contact 50.5%)๊ฐ€ uniform 100๊ฐœ(51.4%)์™€ ๊ฐ™์€ ์„ฑ๊ณต๋ฅ ์„ ๋‚ด๊ณ , uniform 10ยท15๊ฐœ์™€ Least-contact๋Š” ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ๊ณผ ๊ฐ™์€ 0.0%๋‹ค. 2000๊ฐœ๋ฅผ ๋‹ค ์“ฐ๋ฉด layout์€ CNN์—์„œ๋„ ์„ฑ๋Šฅ์„ ๋ฐ”๊พธ์ง€ ์•Š๋Š”๋‹ค(61.3~64.3%). ๋‹ค๋งŒ Grasp-and-Lift(SharpaยทAllegro)์—์„œ๋Š” ๊ฒฝํ–ฅ์ด ๋‹ฌ๋ผ์ ธ Top-contact๊ฐ€ uniform์„ ์ด๊ธฐ์ง€ ๋ชปํ•œ๋‹ค.

๐Ÿ” Ping Review

๐Ÿ” Ping โ€” A light tap on the surface. Get the gist in seconds.

์ด‰๊ฐ policy๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์ „์— ์ด๋ฏธ ์ •ํ•ด์ง€๋Š” ๊ฒƒ๋“ค์ด ์žˆ๋‹ค. ์–ด๋–ค ์„ผ์„œ๋ฅผ ๊ด€์ธก์— ๋„ฃ์„์ง€, ๊ทธ ์„ผ์„œ๋ฅผ ์†์˜ ์–ด๋””์— ๋‘˜์ง€, ๊ทธ ๊ฐ’์„ encoder๊ฐ€ ์ฝ์„ tensor์— ์–ด๋–ค ์ˆœ์„œ๋กœ ๋„ฃ์„์ง€๋‹ค. ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์—ฐ๊ตฌ๋Š” โ€œ์ •์ฑ…์ด ์–ด๋””๋ฅผ ๋งŒ์งˆ์ง€ ๋ชจ๋ฅธ๋‹คโ€๋Š” ์ด์œ ๋กœ taxel์„ ์† ์ „์ฒด์— uniformํ•˜๊ฒŒ ๊น”์•„ ์™”๊ณ , ์ด ๊ด€ํ–‰์ด ๊ฒ€์ฆ ์—†์ด ๊ธฐ๋ณธ๊ฐ’์ด ๋๋‹ค. ์ €์ž๋“ค์€ ์ด ๊ธฐ๋ณธ๊ฐ’์„ ํ•˜๋‚˜์˜ ํ†ต์ œ๋œ ํ”„๋กœํ† ์ฝœ ์•ˆ์—์„œ ์‹œํ—˜ํ•œ๋‹ค. teacher, ํ•™์Šต ์ ˆ์ฐจ, ํ›„๋ณด taxel ์ง‘ํ•ฉ(๋งํฌ ํ‘œ๋ฉด์— Poisson-disk sampling์œผ๋กœ ๋ฟŒ๋ฆฐ N=2000๊ฐœ)์€ ๋ชจ๋‘ ๊ณ ์ •ํ•˜๊ณ , taxel์„ ๊ณ ๋ฅด๋Š” ๊ทœ์น™๊ณผ ๊ฐœ์ˆ˜, ๊ทธ๋ฆฌ๊ณ  ํ‘œํ˜„๋งŒ ๋ฐ”๊พผ๋‹ค.

๊ธฐ์กด ๋น„๊ต์™€ ๊ฒฐ์ •์ ์œผ๋กœ ๋‹ค๋ฅธ ์ ์€ ๋น„๊ต ๋‹จ์œ„๋‹ค. Tactile Genesis๋Š” fingertip ๋ฐฐ์น˜์™€ ์† ์ „์ฒด ๋ฐฐ์น˜๋ฅผ ๋น„๊ตํ–ˆ์ง€๋งŒ ๋‘ ์กฐ๊ฑด์€ taxel ์ˆ˜๋ถ€ํ„ฐ ๋‹ฌ๋ž๋‹ค. ์ด ๋…ผ๋ฌธ์€ ํ›„๋ณด ์ง‘ํ•ฉ๊ณผ taxel ์ˆ˜ B๋ฅผ ๊ฐ™๊ฒŒ ๋‘๊ณ  ์„ ํƒ ๊ทœ์น™๋งŒ ๋‹ค๋ฅด๊ฒŒ ํ•œ๋‹ค. ๊ทธ๋ž˜์„œ โ€œ๋ช‡ ๊ฐœ๋ƒโ€์™€ โ€œ์–ด๋А ๊ฒƒ์ด๋ƒโ€๊ฐ€ ์ฒ˜์Œ์œผ๋กœ ๋ถ„๋ฆฌ๋œ๋‹ค.


Fig. 2. XHand1 In-Palm Rotation์˜ taxel placement. ์™ผ์ชฝ์€ teacher ฯ€_T๊ฐ€ ๊ฐ ํ›„๋ณด taxel์—์„œ ์ ‘์ด‰ํ•œ ๋นˆ๋„(์ƒ‰์ด ์ง„ํ• ์ˆ˜๋ก ์ž์ฃผ ์ ‘์ด‰, ํšŒ์ƒ‰์€ 10โปโด ๋ฏธ๋งŒ), ์˜ค๋ฅธ์ชฝ์€ B=100์ผ ๋•Œ ๋„ค ๊ฐ€์ง€ ๊ทœ์น™์ด ๊ณ ๋ฅธ taxel์ด๋‹ค. ์ด ๊ณผ์ œ์—์„œ๋Š” ์—„์ง€ยท์ค‘์ง€ยท์†๋ฐ”๋‹ฅ๋งŒ ์“ฐ๊ณ , ํ๋ฆฌ๊ฒŒ ๊ทธ๋ฆฐ ์†๊ฐ€๋ฝ์€ ์“ฐ์ง€ ์•Š๋Š”๋‹ค.

ํ•ต์‹ฌ ๋ฐฉ๋ฒ•๋ก :

๊ฐ ๊ณผ์ œ๋Š” POMDP๋กœ ๋‘๊ณ  ๋‘ ๋‹จ๊ณ„๋กœ ํ•™์Šตํ•œ๋‹ค. ๋จผ์ € ๊ด€์ธกํ•  ์ˆ˜ ์—†๋Š” ๋ฌผ์ฒด ์†์„ฑ(privileged)์„ ๋ฐ›๋Š” teacher \pi_T๋ฅผ PPO๋กœ ํ•™์Šตํ•œ๋‹ค. teacher๋Š” ์ด‰๊ฐ์„ ๋ฐ›์ง€ ์•Š์œผ๋ฏ€๋กœ ๊ทธ ์ ‘์ด‰ ์œ„์น˜๋Š” taxel placement์™€ ๋ฌด๊ด€ํ•˜๋‹ค. ๋‹ค์Œ์œผ๋กœ ๋ฐฐํฌ ๋•Œ ์“ธ ์ˆ˜ ์žˆ๋Š” ๊ด€์ธก(๊ด€์ ˆ ์ƒํƒœ ์ด๋ ฅ, ๋ชฉํ‘œ, ์ด‰๊ฐ o^{\mathrm{tac}}_t)๋งŒ ๋ฐ›๋Š” ํ•™์ƒ \pi_S๋ฅผ DAgger๋กœ distillํ•œ๋‹ค.

\min_{\pi_S}\; \mathbb{E}\left[\left\lVert \pi_S(o_{1:t}) - \pi_T(s_t) \right\rVert_2^2\right]

๊ธฐ๋Œ“๊ฐ’์€ \pi_S๊ฐ€ ๋ฐฉ๋ฌธํ•œ ์ƒํƒœ์— ๋Œ€ํ•ด ์žก๋Š”๋‹ค. ๋ชจ๋“  ํ•™์ƒ์€ ๊ฐ™์€ \pi_T์—์„œ distill๋˜๊ณ  ์ด‰๊ฐ ๊ด€์ธก๊ณผ ๊ทธ encoder๋งŒ ๋‹ค๋ฅด๋‹ค.

taxel i๋Š” ์ž๊ธฐ ๋งํฌ ์ขŒํ‘œ๊ณ„์˜ ํž˜ยทํ† ํฌ x_{i,t}\in\mathbb{R}^6๋ฅผ ์žฐ๋‹ค. \mathcal{E}_i๋ฅผ teacher rollout์—์„œ taxel i๊ฐ€ 0์ด ์•„๋‹Œ ๊ฐ’์„ ๋‚ธ timestep ์ง‘ํ•ฉ์ด๋ผ ํ•˜๋ฉด, ์„ ํƒ ๊ทœ์น™์€ ์ด ํ†ต๊ณ„๋ฅผ ์–ด๋–ป๊ฒŒ ์“ฐ๋А๋ƒ๋กœ ๊ฐˆ๋ฆฐ๋‹ค. Top-contact๋Š” |\mathcal{E}_i|๊ฐ€ ํฐ ์ˆœ์„œ๋กœ B๊ฐœ, Contact-cover๋Š” ๊ฐ™์€ ๋งํฌ์—์„œ ์•„์ง ์žกํžˆ์ง€ ์•Š์€ ์ ‘์ด‰ ์ด๋ฒคํŠธ๋ฅผ ๊ฐ€์žฅ ๋งŽ์ด ์žก๋Š” taxel์„ greedyํ•˜๊ฒŒ, Link-proportional์€ ๋งํฌ๋ณ„ ์ ‘์ด‰ ๋น„์ค‘๋Œ€๋กœ ๊ฐœ์ˆ˜๋ฅผ ๋‚˜๋ˆˆ๋‹ค. Uniform์€ ๊ณผ์ œ ์ •๋ณด ์—†์ด ๋งํฌ๋ณ„ ํ›„๋ณด ์ˆ˜์— ๋น„๋ก€ํ•ด ๋‚˜๋ˆ„๊ณ , Least-contact๋Š” ์ ‘์ด‰์€ ๋์ง€๋งŒ |\mathcal{E}_i|๊ฐ€ ๊ฐ€์žฅ ์ž‘์€ taxel์„ ๊ณ ๋ฅด๋Š” ๋Œ€์กฐ๊ตฐ์ด๋‹ค.

์ฃผ์š” ๊ฒฐ๊ณผ:

  • In-Palm Rotation, XHand1, LSTM, 3 seeds (Table 1): Top-contact B=10์ด 41.4%์ธ๋ฐ Uniform์€ B=10ยท15์—์„œ 0.0%, B=100์—์„œ์•ผ 51.4%๋‹ค. Top-contact B=25(50.5%)๊ฐ€ Uniform B=100๊ณผ ๊ฐ™๊ณ , Top-contact B=75(59.5%)๋„ Uniform B=300(56.2%)์— ๋’ค์ง€์ง€ ์•Š๋Š”๋‹ค. 2000๊ฐœ ์ „๋ถ€๋Š” 61.5%, teacher๋Š” 93.8%๋‹ค.
  • ์ด‰๊ฐ์ด ์—†๋Š” ๊ฒƒ๊ณผ ๊ฐ™์€ taxel: Least-contact(B=50ยท100)๋Š” 0.0%๋กœ, ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ(0.0%)๊ณผ ๊ฐ™๋‹ค. ๊ณต๊ฐ„ ํ•ด์ƒ๋„ ์—†๋Š” Per-link force๋Š” 4.4%๋‹ค.
  • Grasp-and-Lift, SharpaยทAllegro (Table 2): ๊ฒฝํ–ฅ์ด ๋ฐ˜๋Œ€๋‹ค. ๋‚ฎ์€ B์—์„œ๋Š” ๋ฐฉ๋ฒ•๋ผ๋ฆฌ ์ฐจ์ด๊ฐ€ ์ž‘๊ณ  ๊ฐ€์žฅ ํฐ B=300์—์„œ ๊ฐˆ๋ฆฐ๋‹ค. Contact-cover๊ฐ€ ๋‘ ์† ๋ชจ๋‘์—์„œ Uniform๋ณด๋‹ค ๋†’๊ณ (Sharpa 57.0 vs 42.6, Allegro 72.2 vs 57.0), Top-contact๋Š” Sharpa B=300์—์„œ 27.0%๋กœ Uniform๋ณด๋‹ค ํฌ๊ฒŒ ๋‚ฎ์œผ๋ฉฐ Allegro B=25์—์„œ๋Š” ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ๊ณผ ๊ฐ™๋‹ค(22.8 vs 22.9).
  • ํ‘œํ˜„ (Table 3): 2000๊ฐœ๋ฅผ ๋‹ค ์“ฐ๋ฉด CNN์€ Random 61.4 / Band 64.3 / Grid 61.3, LSTM์€ 60.8 / 58.0 / 61.5๋กœ layout ์ฐจ์ด๊ฐ€ ์ž‘๋‹ค. taxel์˜ ์ •์ฒด์„ฑ(identity)์„ ์ง€์šฐ๋Š” Link-wise MLP๋งŒ 18.6~33.2%๋กœ ํฌ๊ฒŒ ๋–จ์–ด์ง„๋‹ค.
  • ๋น„์šฉ (Table 4, H100, 2048 env): ์„ฑ๊ณต๋ฅ ์ด ํฌํ™”ํ•˜๋Š” B=100๋งŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋ฉด 29,503 env-steps/s๋กœ, ์ด ๋…ผ๋ฌธ์ด ์‹ค์ œ๋กœ ํ•™์Šต์— ์“ด full observation(5,610)๋ณด๋‹ค 5.3๋ฐฐ ๋น ๋ฅด๋‹ค.

๊ฒฐ๋ก :

์ด ๋…ผ๋ฌธ์˜ ๋ฉ”์‹œ์ง€๋Š” โ€œ์ด‰๊ฐ ์‹ ํ˜ธ๋Š” ์ ‘์ด‰์ด ์ผ์–ด๋‚˜๋Š” ๊ณณ์—์„œ๋งŒ ์ •๋ณด๋ฅผ ์ค€๋‹คโ€๋Š” ํ•œ ๋ฌธ์žฅ์œผ๋กœ ๋ชจ์ธ๋‹ค. uniform placement๋Š” ์ด ์„ฑ์งˆ์„ ๋ฌด์‹œํ•˜๋ฏ€๋กœ, ๋Œ€๋ถ€๋ถ„์˜ taxel์ด 0์„ ๋‚ด๋Š” ๋™์•ˆ taxel ์ˆ˜๋งŒ ๋Š˜๋ ค ๊ณผ์ œ๊ฐ€ ์‹ค์ œ๋กœ ์š”๊ตฌํ•˜๋Š” ์ˆ˜๋ฅผ ๊ณผ๋Œ€ํ‰๊ฐ€ํ•œ๋‹ค. ๊ทธ ์ฆ๊ฑฐ๋Š” XHand1 In-Palm Rotation ํ•œ ๊ณผ์ œยทํ•œ ์†์—์„œ ๊ฐ€์žฅ ๊นจ๋—ํ•˜๊ณ , ์ €์ž ์Šค์Šค๋กœ โ€œ์ ‘์ด‰์ด ๋ช‡ ๊ฐœ ๋งํฌ์— ๋ชฐ๋ฆฌ๋Š” ํ•œ ์†ยทํ•œ ๊ณผ์ œ์—์„œ ๋‚˜์˜จ ๊ตํ›ˆโ€์ด๋ผ๊ณ  ๋ฒ”์œ„๋ฅผ ๊ธ‹๋Š”๋‹ค. Grasp-and-Lift์—์„œ๋Š” Top-contact๊ฐ€ ์‹คํŒจํ•˜๊ณ  โ€œ์ ‘์ด‰ํ•œ ๋งํฌ๋ฅผ ๋น ์ง์—†์ด ๋ฎ๋Š” ๊ฒƒโ€์ด ๋” ์ค‘์š”ํ•ด์ง„๋‹ค. ๊ทธ๋ž˜์„œ ์‹ค๋ฌด ๋ฉ”์‹œ์ง€๋Š” โ€œ์ž์ฃผ ๋‹ฟ๋Š” ๊ณณ์— ๋ชฐ์•„๋ผโ€๊ฐ€ ์•„๋‹ˆ๋ผ โ€œteacher ์ ‘์ด‰ ํ†ต๊ณ„๋ฅผ ๋จผ์ € ๋ณด๊ณ , ๋‹ฟ๋Š” ๋งํฌ๋ฅผ ๋ฎ๋„๋ก ๊ณจ๋ผ๋ผโ€์— ๊ฐ€๊น๋‹ค. ๋ถ€๋ก์—๋Š” ์ด‰๊ฐ policy ์—ฐ๊ตฌ๊ฐ€ ๋ณด๊ณ ํ•ด์•ผ ํ•  ์„ค๊ณ„ ์„ ํƒ์˜ checklist(Table 5)๊ฐ€ ๋ถ™์–ด ์žˆ๋‹ค.


๐Ÿ”” Ring Review

๐Ÿ”” Ring โ€” An idea that echoes. Grasp the core and its value.

ํ•œ ์ค„๋กœ ์‹œ์ž‘ํ•˜๋ฉด

๊ฐ™์€ teacher, ๊ฐ™์€ ํ•™์Šต ์ ˆ์ฐจ, ๊ฐ™์€ ํ›„๋ณด taxel 2000๊ฐœ๋ฅผ ๋‘๊ณ  โ€œ์–ด๋–ค B๊ฐœ๋ฅผ ๊ณ ๋ฅด๋А๋ƒโ€๋งŒ ๋ฐ”๊ฟจ๋”๋‹ˆ, uniform์œผ๋กœ 100๊ฐœ๋ฅผ ๊น”์•„์•ผ ์–ป๋Š” ์„ฑ๊ณต๋ฅ ์„ ์ ‘์ด‰ ํ†ต๊ณ„๋กœ ๊ณ ๋ฅธ 25๊ฐœ๊ฐ€ ๋ƒˆ๋‹ค. ๊ทธ ์ฐจ์ด๋Š” ์ •๋ณด๋Ÿ‰์ด ์•„๋‹ˆ๋ผ 0์˜ ์–‘์—์„œ ์˜จ๋‹ค. uniform์œผ๋กœ ๊ณ ๋ฅธ taxel ๋Œ€๋ถ€๋ถ„์€ teacher๊ฐ€ ํ•œ ๋ฒˆ๋„ ๋‹ฟ์ง€ ์•Š๋Š” ๊ณณ์— ์žˆ์–ด 0๋งŒ ๋‚ด๊ณ , ๊ทธ๋ž˜์„œ uniform์€ โ€œ๋งŽ์ด ํ•„์š”ํ•œ ๊ฒƒ์ฒ˜๋Ÿผโ€ ๋ณด์ธ๋‹ค. ์ด ๋ฆฌ๋ทฐ๋Š” ๊ทธ ์ฃผ์žฅ์„ ์ธ์ •ํ•˜๋˜, ๊ทธ๊ฒƒ์ด ์–ด๋А ์กฐ๊ฑด(XHand1ยทIn-Palm RotationยทLSTMยท3 seeds)์—์„œ ๊นจ๋—ํ•˜๊ฒŒ ์„ฑ๋ฆฝํ•˜๊ณ  ์–ด๋А ์กฐ๊ฑด(Grasp-and-Lift)์—์„œ ํ๋ ค์ง€๋Š”์ง€, ๊ทธ๋ฆฌ๊ณ  ๋น„๊ต ์„ค๊ณ„์— ์ˆจ์€ ๋น„๋Œ€์นญ์ด ๊ฒฐ๋ก ์„ ์–ผ๋งˆ๋‚˜ ํ”๋“œ๋Š”์ง€๋ฅผ ๋…ผ๋ฌธ์˜ ํ‘œ๋กœ ๋”ฐ์ ธ ๋ณธ๋‹ค.

๋ฐฐ๊ฒฝ: placement๋Š” ์™œ ์ˆจ์€ ๊ธฐ๋ณธ๊ฐ’์ด ๋๋‚˜ (ยง1ยทยง2)

์ €์ž๋“ค์€ ์ด‰๊ฐ ์‹ ํ˜ธ๊ฐ€ ์‹œ๊ฐยทproprioception๊ณผ ์„ฑ์งˆ์ด ๋‹ค๋ฅด๋‹ค๋Š” ๋ฐ์„œ ์ถœ๋ฐœํ•œ๋‹ค.

  • sparseํ•˜๊ณ  non-smoothํ•˜๋‹ค. ์ ‘์ด‰์ด ์žˆ์„ ๋•Œ๋งŒ ๊ฐ’์ด ์ƒ๊ธด๋‹ค. ์ด๊ฒƒ์ด ๋ถˆ๊ท ํ˜•ํ•œ ํ•™์Šต ๋ฐ์ดํ„ฐ์™€ ๋ถˆ์•ˆ์ •ํ•œ ํ•™์Šต์œผ๋กœ ์ด์–ด์งˆ ์ˆ˜ ์žˆ๋‹ค(Miller et al., 2025 ์ธ์šฉ).
  • ๊ณต๊ฐ„์ ์œผ๋กœ ๋ถˆ๊ทœ์น™ํ•˜๋‹ค. ์†์˜ ๊ด€์ ˆ ๊ตฌ์กฐ ๋•Œ๋ฌธ์— ์ ‘์ด‰์€ ์˜์—ญ๋งˆ๋‹ค ๊ณ ๋ฅด์ง€ ์•Š๊ฒŒ ๋ถ„ํฌํ•˜๊ณ , taxel์€ ๊ทœ์น™์ ์ธ ๊ฒฉ์ž๊ฐ€ ์•„๋‹ˆ๋ผ ๊ณก๋ฉด ์œ„์— ๋†“์ธ๋‹ค.
  • ํ•„์š”ํ•œ ๊ณต๊ฐ„ ํ•ด์ƒ๋„๊ฐ€ ๊ณผ์ œ๋งˆ๋‹ค ๋‹ค๋ฅด๋‹ค. antipodal grasping์—์„œ๋Š” ๊ณต๊ฐ„ ๋ถ„ํ•ด๋œ ์ด‰๊ฐ์ด ์†๊ฐ€๋ฝ๋ณ„ net force ์ด์ƒ์„ ์ฃผ์ง€ ๋ชปํ•˜๊ณ (Zhang et al., 2025), in-hand rotation์—์„œ๋Š” ์ ‘์ด‰ ์œ„์น˜๊ฐ€ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ์˜ฌ๋ฆฐ๋‹ค(RotateIt, Qi et al., 2023).

์ •์ฑ…์ด ์–ด๋””๋ฅผ ๋งŒ์งˆ์ง€๋Š” ํ•™์Šต ์ „์— ์•Œ ์ˆ˜ ์—†์œผ๋ฏ€๋กœ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์—ฐ๊ตฌ๋Š” taxel์„ uniformํ•˜๊ฒŒ ๊น”๊ณ , ์ด ๋…ผ๋ฌธ์˜ ํ›„๋ณด ์ง‘ํ•ฉ๋„ ๊ฐ™์€ ๊ด€ํ–‰์„ ๋”ฐ๋ฅธ๋‹ค. ์ €์ž๋“ค์€ ์ด๋ฅผ โ€œ์ธก์ •๋œ ์š”๊ตฌ์กฐ๊ฑด์ด ์•„๋‹ˆ๋ผ ๊ฐ€์ •โ€์ด๋ผ๊ณ  ๋ถ€๋ฅธ๋‹ค. ํ•œ ๋ฒˆ๋„ ๋‹ฟ์ง€ ์•Š๋Š” taxel์€ 0๋งŒ ๋‚ด์ง€๋งŒ ๊ด€์ธก ์ฐจ์›๊ณผ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ์‹œ๊ฐ„์„ ์“ฐ๊ณ , ์‹ค์ œ ์†์—์„œ๋Š” ๋ฐฐ์„ ๊ณผ readout ๋Œ€์—ญํญ์„ ์“ด๋‹ค.

๊ด€๋ จ ์—ฐ๊ตฌ ์ ˆ(ยง2)์€ ๋‘ ๊ฐˆ๋ž˜๋‹ค.

  • ํ‘œํ˜„(ยง2.1). sequence model์€ ์ˆœ์„œ ์žˆ๋Š” token์„, CNN์€ ๊ทœ์น™์ ์ธ pixel ๊ฒฉ์ž๋ฅผ ๊ฐ€์ •ํ•˜์ง€๋งŒ, ๊ด€์ ˆ ๋งํฌ์˜ ๊ณก๋ฉด ์œ„ taxel์€ ๋‘˜ ๋‹ค ๋งŒ์กฑํ•˜์ง€ ์•Š๋Š”๋‹ค. ๊ธฐ์กด dexterous policy๋Š” ๊ฒฉ์ž ์ž…๋ ฅ์— CNN์„, ๊ฒฉ์ž ์—†๋Š” ์ž…๋ ฅ์— MLP๋‚˜ LSTM์„ ๋ถ™์—ฌ ์™”๊ณ  GNN๋„ ์“ฐ์˜€๋‹ค(TacGNN). Tactile Genesis๋Š” tensor layout์„ ๊ณ ์ •ํ•˜๊ณ  ๊ด€์ธก ์ข…๋ฅ˜๋ฅผ ๋น„๊ตํ•ด per-taxel force/torque๊ฐ€ ๊ฐ€์žฅ ์ผ๊ด€๋˜๊ฒŒ ์œ ์šฉํ•˜๋‹ค๊ณ  ํ–ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ๋ฐ˜๋Œ€๋กœ ๊ทธ ์‹ ํ˜ธ๋ฅผ ๊ณ ์ •ํ•˜๊ณ  layout๊ณผ encoder๋ฅผ ๋ถ„๋ฆฌํ•œ๋‹ค.
  • ์„ผ์„œ ๊ตฌ์„ฑ(ยง2.2). ํ•ด์ƒ๋„๊ฐ€ ํ•™์Šต ํšจ์œจ์— ์ฃผ๋Š” ์˜ํ–ฅ(Melnik et al., 2021), ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•˜๋ฉฐ ์„ผ์„œ ์ˆ˜๋ฅผ ์ค„์ด๊ธฐ(Guo et al., 2024), ์†๋ฐ”๋‹ฅยท์†๊ฐ€๋ฝ ์˜์—ญ๋ณ„ ๊ธฐ์—ฌ(RoleOfTouch, Almeida et al., 2025), ๊ฐ™์€ ์„ผ์„œ ์ˆ˜์—์„œ๋„ placement๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”๊พธ๊ณ  ๊ทธ ํšจ๊ณผ๊ฐ€ ์† ๋ชจ๋ธยทํ•™์Šต ์•Œ๊ณ ๋ฆฌ์ฆ˜์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง„๋‹ค๋Š” ๋ณด๊ณ (Birtalan & Koller, 2026)๊ฐ€ ์žˆ๋‹ค. Tactile Genesis์˜ โ€œ์† ์ „์ฒด๊ฐ€ fingertip๋ณด๋‹ค ๋‚ซ๋‹คโ€๋Š” ๊ฒฐ๊ณผ๋Š” ๋‘ ์กฐ๊ฑด์˜ ์„ผ์„œ ์ˆ˜๊ฐ€ ๋‹ค๋ฅด๊ณ  taxel๋„ uniformํ•˜๊ฒŒ ๋ฝ‘์•˜๋‹ค๋Š” ์ ์—์„œ, ์ด ๋…ผ๋ฌธ์ด ๋ฉ”์šฐ๋ ค๋Š” ๋นˆ์นธ์„ ์ •ํ™•ํžˆ ๋ณด์—ฌ ์ค€๋‹ค.

์—ฐ๊ตฌ ์งˆ๋ฌธ (ยง3.1)

  • RQ1 (Sensor configuration): taxel์˜ ๊ฐœ์ˆ˜์™€ placement๊ฐ€ policy ์„ฑ๋Šฅ์— ์–ด๋–ป๊ฒŒ ์˜ํ–ฅ์„ ์ฃผ๋Š”๊ฐ€, ๊ทธ๋ฆฌ๊ณ  placement์˜ ํšจ๊ณผ๋Š” taxel budget์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์ง€๋Š”๊ฐ€.
  • RQ2 (Representation): ๊ฐ™์€ taxel ์‹ ํ˜ธ๊ฐ€ ์ฃผ์–ด์งˆ ๋•Œ tensor layout๊ณผ encoder๋Š” ์„ฑ๋Šฅ์— ์–ด๋–ป๊ฒŒ ์˜ํ–ฅ์„ ์ฃผ๋Š”๊ฐ€.

๊ณผ์ œ์™€ embodiment (ยง3.2, ๋ถ€๋ก .2)


Fig. 1. ๊ฐ™์€ ์ถ•์ฒ™์œผ๋กœ ๊ทธ๋ฆฐ ์„ธ ์†. XHand1(12 DoF), Sharpa(22 DoF), Allegro(16 DoF)์˜ ์ •๋ฉดยท์ธก๋ฉด.

๋ชจ๋“  ๊ณผ์ œ๋Š” Genesis(genesis-world)์—์„œ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•œ๋‹ค.

  • In-Palm Rotation (XHand1). ์†๋ฐ”๋‹ฅ ์œ„์— ๋ฌผ์ฒด๋ฅผ ์œ ์ง€ํ•˜๋ฉด์„œ ๊ณ ์ •๋œ ์ˆ˜ํ‰์ถ•์„ ์ค‘์‹ฌ์œผ๋กœ ํ•œ ๋ฐฉํ–ฅ์œผ๋กœ 90๋„์”ฉ ์—ฐ์† ํšŒ์ „์‹œํ‚จ๋‹ค. ์—„์ง€ยท์ค‘์ง€ยท์†๋ฐ”๋‹ฅ๋งŒ ์“ด๋‹ค. ๊ณผ์ œ ์ •์˜๋Š” Tactile Genesis๋ฅผ ๋”ฐ๋ฅธ๋‹ค.
  • Grasp-and-Lift (Sharpa, Allegro). ํ…Œ์ด๋ธ” ์œ„ ๋ฌผ์ฒด๋ฅผ ์žก์•„ ๋“ค์–ด ์˜ฌ๋ ค ํ…Œ์ด๋ธ” ์œ„ ๋ชฉํ‘œ ์œ„์น˜๋กœ ์˜ฎ๊ธด๋‹ค. ๊ณต์ค‘์— ์žˆ๋Š” ๋™์•ˆ ์™ธ๋ž€์„ ๊ฐ€ํ•˜๊ณ  policy๋Š” ํŒŒ์ง€๋ฅผ ์œ ์ง€ํ•ด์•ผ ํ•œ๋‹ค.

์„ธ ์†์€ ํ•˜๋‚˜์˜ ๋ฌผ์ฒด ์ง‘ํ•ฉ(์ƒ์ž 15๊ฐœ, ์›๊ธฐ๋‘ฅ 2๊ฐœ)์„ ์† ํฌ๊ธฐ์— ๋งž์ถฐ ์ถ•์ฒ™ํ•ด ๊ณต์œ ํ•œ๋‹ค. ๋ณ‘๋ ฌ ํ™˜๊ฒฝ๋งˆ๋‹ค ๋ฌผ์ฒด ํ•˜๋‚˜๋ฅผ ์žฅ๋ฉด ์ƒ์„ฑ ๋•Œ ๋ฐฐ์ •ํ•ด ์—ํ”ผ์†Œ๋“œ ๋‚ด๋‚ด ๊ณ ์ •ํ•˜๊ณ , ๋ฌผ์ฒด๋งˆ๋‹ค ํ™˜๊ฒฝ ์ˆ˜๋ฅผ ๊ฐ™๊ฒŒ ๋‚˜๋ˆˆ๋‹ค. ๋ฌผ์ฒด ์งˆ๋Ÿ‰ยท๋งˆ์ฐฐ, ์ดˆ๊ธฐ ์ž์„ธ, actuator ํŒŒ๋ผ๋ฏธํ„ฐ๋Š” ๋งค reset๋งˆ๋‹ค randomizeํ•˜๊ณ  ํ•™์ƒ๋„ teacher์™€ ๊ฐ™์€ randomization์„ ์“ด๋‹ค. In-Palm Rotation์€ ๋ฌผ์ฒด๊ฐ€ ์†์—์„œ ๋–จ์–ด์ง€๋ฉด, Grasp-and-Lift๋Š” ๋ฌผ์ฒด๊ฐ€ ํ…Œ์ด๋ธ” ์•„๋ž˜๋กœ ๋–จ์–ด์ง€๋ฉด ์—ํ”ผ์†Œ๋“œ๊ฐ€ ๋๋‚œ๋‹ค.


Fig. 4. ๋‘ ๊ณผ์ œ์˜ ๋ฌผ์ฒด(์ƒ์ž 15๊ฐœ, ์›๊ธฐ๋‘ฅ 2๊ฐœ ์ค‘ 5๊ฐœ๋ฅผ Allegro ์ถ•์ฒ™์œผ๋กœ ํ‘œ์‹œ, ๋‹จ์œ„ mm).

Fig. 5. ์„ธ ์†์˜ teacher rollout. ๊ฐ ํ–‰์€ ฯ€_T ํ•œ ์—ํ”ผ์†Œ๋“œ์—์„œ 2์ดˆ ๊ฐ„๊ฒฉ์œผ๋กœ ๋ฝ‘์€ ํ”„๋ ˆ์ž„์ด๋‹ค. ์œ„ ๋‘ ํ–‰์€ AllegroยทSharpa์˜ Grasp-and-Lift, ์•„๋ž˜ ํ–‰์€ XHand1์˜ In-Palm Rotation.

๋ฐฉ๋ฒ• ์ƒ์„ธ 1: teacherโ€“student distillation (ยง3.3, ๋ถ€๋ก .3)

teacher \pi_T:\mathcal{S}\to\mathcal{A}๋Š” ๊ด€์ธกํ•  ์ˆ˜ ์—†๋Š” ๋ฌผ์ฒด ์†์„ฑ์„ ๋ฐ›์•„ PPO๋กœ ํ•™์Šตํ•˜๊ณ  ๊ทธ ๋’ค ๊ณ ์ •ํ•œ๋‹ค. ์ด ์„ค๊ณ„์˜ ํ•ต์‹ฌ์€ teacher๊ฐ€ ์ด‰๊ฐ์„ ๋ฐ›์ง€ ์•Š๋Š”๋‹ค๋Š” ์ ์ด๋‹ค. ๊ทธ๋ž˜์„œ teacher์˜ ์ ‘์ด‰ ์œ„์น˜๋Š” placement์™€ ๋ฌด๊ด€ํ•˜๊ณ , placement ๊ทœ์น™๋“ค์ด ๊ฐ™์€ ์ ‘์ด‰ ํ†ต๊ณ„๋ฅผ ๊ณต์œ ํ•  ์ˆ˜ ์žˆ๋‹ค. ํ•™์ƒ \pi_S๋Š” ๊ด€์ ˆ ์ƒํƒœ ์ด๋ ฅ, ๋ชฉํ‘œ, ์ด‰๊ฐ ๊ด€์ธก๋งŒ ๋ฐ›๊ณ  DAgger(\beta=0)๋กœ executed-action MSE๋ฅผ ์ตœ์†Œํ™”ํ•œ๋‹ค(Eq. 1).

๋ณด์ƒ์€ ๊ณผ์ œ ๋ณด์ƒ๊ณผ ์ •๊ทœํ™” ๋ณด์ƒ์˜ ํ•ฉ \mathcal{R}=\mathcal{R}_{\text{task}}+\mathcal{R}_{\text{reg}}์ด๊ณ , ๊ฐ๊ฐ ํ•ญ๋ชฉ์˜ ๊ฐ€์ค‘ํ•ฉ์ด๋‹ค. ๋ณด์ƒ์€ teacher ํ•™์Šต๊ณผ ํ•™์ƒ ํ•™์Šต์˜ ๋ชจ๋‹ˆํ„ฐ๋ง์—๋งŒ ์“ด๋‹ค.

ํ•ญ๋ชฉ In-Palm Rotation (Table 6) ๊ฐ€์ค‘์น˜
Goal Reached \mathbb{1}_{\theta_t<0.1} 150
Rotation Progress \min(\Delta\theta_t, 6) 2
Fingertip Proximity \exp(-d_{\min,t}^2/0.04^2) 0.1
Drop \mathbb{1}_{\text{dropped}} โˆ’100
Off-Axis Orientation \theta^{\perp}_t โˆ’1
Object Displacement [\lVert p_t-p^*\rVert_2-0.02]_+ โˆ’100
Contact Force \min([\lVert f_t\rVert_2-1]_+, 20) โˆ’5
Action Rate \lVert a_t-a_{t-1}\rVert_2^2 -5\times10^{-4}
Joint Limits ๊ด€์ ˆ ํ•œ๊ณ„ ์ดˆ๊ณผ๋Ÿ‰์˜ ํ•ฉ โˆ’40
Work \sum_j \lvert\tau_{j,t}\dot q_{j,t}\rvert -5\times10^{-3}
ํ•ญ๋ชฉ Grasp-and-Lift (Table 7, SharpaยทAllegro ๊ณตํ†ต) ๊ฐ€์ค‘์น˜
Sustained Lift h_t + 10\,\mathbb{1}_{h_t>0.05} 5
Goal Tracking \exp(-\lVert p_t-p^{\text{goal}}\rVert_2^2/0.25^2) 1
Drop \mathbb{1}_{\text{dropped}} โˆ’30
Fingertip Proximity fingertip-๋ฌผ์ฒด ๊ฑฐ๋ฆฌ์˜ \exp(-d^2/0.03^2) ํ‰๊ท  1
Palm Orientation [-\hat n_t\cdot\hat z]_+ (์†๋ฐ”๋‹ฅ์ด ์•„๋ž˜๋ฅผ ํ–ฅํ•˜๋ฉด ๋ฒŒ์ ) โˆ’0.5
Action Rate \lVert a_t-a_{t-1}\rVert_2^2 -10^{-3}

ํ•™์Šต ์„ค์ •(Table 8, ๊ฐ’์ด ๋‘˜์ด๋ฉด In-Palm Rotation / Grasp-and-Lift):

๊ตฌ๋ถ„ ์„ค์ • ๊ฐ’
Simulation physics / control step 5 ms / 25 ms
episode length 10 s / 20 s
Teacher (PPO) parallel envs 8192 / 4096
steps per env, iterations 24, 20000 / 5000
epochs, mini-batches 5, 4
learning rate 10^{-3}, adaptive (target KL 0.01)
\gamma, \lambda, clip 0.998, 0.95, 0.2
entropy coef., grad clip 0.002 / 0.001, 1.0
actor, critic MLP 512-256-128, ELU
Student (DAgger) scheme, loss DAgger \beta=0, clipped action MSE
parallel envs, steps per env, iterations 2048, 24, 3000
auxiliary losses object size, goal angle (MSE, weight 1.0) / none
optimizer Adam 10^{-4}, 5 epochs
policy head MLP 512-256-128, ELU, fixed std 1.0

In-Palm Rotation ํ•™์ƒ์—๋งŒ ๋ฌผ์ฒด ํฌ๊ธฐยท๋ชฉํ‘œ ๊ฐ๋„๋ฅผ ์˜ˆ์ธกํ•˜๋Š” auxiliary loss๊ฐ€ ๋ถ™๋Š”๋‹ค๋Š” ์ ์€ ๊ธฐ์–ตํ•ด ๋‘˜ ๋งŒํ•˜๋‹ค. ํ•™์ƒ์€ ๋ฌผ์ฒด pose๋ฅผ ๊ด€์ธกํ•˜์ง€ ์•Š์œผ๋ฏ€๋กœ, ์ด ๊ณผ์ œ์˜ ์ด‰๊ฐ์€ โ€œ๋ฌผ์ฒด๊ฐ€ ์ง€๊ธˆ ์–ด๋–ป๊ฒŒ ๋†“์—ฌ ์žˆ๋Š”๊ฐ€โ€๋ฅผ ๋Œ€์‹  ์ถ”๋ก ํ•˜๋Š” ๊ฑฐ์˜ ์œ ์ผํ•œ ๊ฒฝ๋กœ๋‹ค.

๋ฐฉ๋ฒ• ์ƒ์„ธ 2: taxel ํ›„๋ณด์™€ placement ๊ทœ์น™ (ยง3.4, ๋ถ€๋ก .5)

Tactile Genesis๋ฅผ ๋”ฐ๋ผ ๊ณผ์ œ์—์„œ ์“ฐ๋Š” ๋งํฌ ํ‘œ๋ฉด์— Poisson-disk sampling์œผ๋กœ ํ›„๋ณด N=2000๊ฐœ๋ฅผ uniformํ•˜๊ฒŒ ๋ฟŒ๋ฆฐ๋‹ค. ๊ฐ taxel์€ ๋งํฌ ์ขŒํ‘œ๊ณ„์˜ 6์ฐจ์› ํž˜ยทํ† ํฌ x_{i,t}๋ฅผ ๋‚ด๊ณ , placement ๊ทœ์น™์€ ๊ทธ์ค‘ B๊ฐœ(10~300)๋ฅผ ๊ณ ๋ฅธ๋‹ค. ์„ค๊ณ„์—์„œ ๋ˆˆ์—ฌ๊ฒจ๋ณผ ํ†ต์ œ ์žฅ์น˜๊ฐ€ ์„ธ ๊ฐ€์ง€๋‹ค.

  • nested ์„ ํƒ. ๊ฐ™์€ ๊ทœ์น™ ์•ˆ์—์„œ budget์ด ์ปค์ง€๋ฉด taxel์„ ๋”ํ•˜๊ธฐ๋งŒ ํ•œ๋‹ค. B์— ๋”ฐ๋ฅธ ์ฐจ์ด๊ฐ€ ๋‹ค๋ฅธ taxel ์ง‘ํ•ฉ ๋•Œ๋ฌธ์ด ์•„๋‹˜์„ ๋ณด์žฅํ•œ๋‹ค.
  • zero-padding. In-Palm Rotation์—์„œ๋Š” ๋ฝ‘ํžˆ์ง€ ์•Š์€ taxel ์ž๋ฆฌ๋ฅผ 0์œผ๋กœ ์ฑ„์›Œ ์ž…๋ ฅ ์ฐจ์›์„ ๋ชจ๋“  budgetยท๊ทœ์น™์—์„œ ๊ฐ™๊ฒŒ ๋‘”๋‹ค.
  • ๊ฐ™์€ ์ ‘์ด‰ ํ†ต๊ณ„. ๊ณผ์ œ ์ •๋ณด๋ฅผ ์“ฐ๋Š” ๊ทœ์น™์€ ๋ชจ๋‘ ๊ฐ™์€ teacher rollout ํ†ต๊ณ„๋ฅผ ์“ฐ๊ณ , ์ฐจ์ด๋Š” ์„ ํƒ ๊ธฐ์ค€๋ฟ์ด๋‹ค.
๊ทœ์น™ ์ข…๋ฅ˜ ๊ธฐ์ค€ ๋งํฌ ๊ฐ„ ๋ถ„ํฌ
Uniform task-agnostic ๋งํฌ๋ณ„ ํ›„๋ณด ์ˆ˜์— ๋น„๋ก€ํ•ด ๋ฐฐ๋ถ„, ๋งํฌ ์•ˆ์€ farthest point sampling ํ‘œ๋ฉด์  ๋น„๋ก€
Top-contact task-informed \lvert\mathcal{E}_i\rvert ์ƒ์œ„ B๊ฐœ ์ž์ฃผ ๋‹ฟ๋Š” ์ง€์ ์— ์ง‘์ค‘
Contact-cover task-informed ๊ฐ™์€ ๋งํฌ์—์„œ ์•„์ง ์•ˆ ์žกํžŒ ์ ‘์ด‰ ์ด๋ฒคํŠธ๋ฅผ ๊ฐ€์žฅ ๋งŽ์ด ์žก๋Š” taxel์„ greedy ์ถ”๊ฐ€ ๋งํฌ์— ๊ณ ๋ฅด๊ฒŒ ํผ์ง
Link-proportional task-informed ๋งํฌ๋ณ„ ์ ‘์ด‰ ๋น„์ค‘๋Œ€๋กœ ๋ฐฐ๋ถ„, ๋งํฌ ์•ˆ์€ farthest point sampling ๋งํฌ์— ํผ์ง
Least-contact ๋Œ€์กฐ๊ตฐ ์ ‘์ด‰๋œ taxel ์ค‘ \lvert\mathcal{E}_i\rvert ํ•˜์œ„ B๊ฐœ ๋“œ๋ฌผ๊ฒŒ ๋‹ฟ๋Š” ์ง€์ 
Per-link force ์ฐธ์กฐ taxel ๋Œ€์‹  ๋งํฌ๋ณ„ net contact force ๊ณต๊ฐ„ ํ•ด์ƒ๋„ ์—†์Œ

์—ฌ๊ธฐ์„œ โ€œ์ ‘์ด‰ ์ด๋ฒคํŠธ(contact event)โ€๋Š” ์–ด๋–ค ๋งํฌ๊ฐ€ ์ ‘์ด‰ ์ค‘์ธ timestep์ด๋‹ค. Contact-cover๋ฅผ ์‹์œผ๋กœ ๋‹ค์‹œ ์“ฐ๋ฉด, ์„ ํƒ๋œ ์ง‘ํ•ฉ S์— ๋Œ€ํ•ด taxel i(๋งํฌ \ell(i))์˜ ์ด๋“์€ โ€œ๋งํฌ \ell(i)์˜ ์ ‘์ด‰ ์ด๋ฒคํŠธ ์ค‘ i๋Š” ์žก์ง€๋งŒ S ์•ˆ์˜ ๊ฐ™์€ ๋งํฌ taxel์€ ์•„๋ฌด๋„ ๋ชป ์žก์€ ์ด๋ฒคํŠธ ์ˆ˜โ€์ด๊ณ , ์ด๋ฅผ ์ตœ๋Œ€ํ™”ํ•˜๋Š” i๋ฅผ ํ•˜๋‚˜์”ฉ ๋”ํ•œ๋‹ค(ํ•„์ž ํ‘œ๊ธฐ). ์‹ค์ œ ๊ณ„์‚ฐ์€ lazy greedy๋กœ ํ•œ๋‹ค(๋ถ€๋ก .5). coverage ํ•จ์ˆ˜๋Š” taxel์ด ๋Š˜์ˆ˜๋ก ์ถ”๊ฐ€ ์ด๋“์ด ์ค„์–ด๋“ค๋ฏ€๋กœ lazy greedy๊ฐ€ ์ผ๋ฐ˜ greedy์™€ ๊ฐ™์€ ์„ ํƒ์„ ๋‚ธ๋‹ค(ํ•„์ž ์„ค๋ช…).

๋ถ€๋ก .5์˜ ์„ธ๋ถ€ ๊ทœ์น™์€ ๊ฒฐ๊ณผ๋ฅผ ์ฝ์„ ๋•Œ ์ค‘์š”ํ•˜๋‹ค.

  • In-Palm Rotation์—์„œ๋งŒ Top-contactยทContact-coverยทLeast-contact๊ฐ€ ๊ณผ์ œ์— ์“ฐ๋Š” ๋ชจ๋“  ๋งํฌ์— ์ตœ์†Œ 1๊ฐœ์”ฉ taxel์„ ๋ฐฐ์ •ํ•œ๋‹ค.
  • ์ ‘์ด‰ ์ด๋ฒคํŠธ๋Š” 4 timestep๋งˆ๋‹ค ํ‰๊ฐ€ํ•œ๋‹ค.
  • Least-contact๋Š” \lvert\mathcal{E}_i\rvert=0์ธ ํ›„๋ณด 879๊ฐœ๋ฅผ ๋บ€๋‹ค. ์ฆ‰ XHand1 In-Palm Rotation์—์„œ๋Š” ํ›„๋ณด์˜ ์•ฝ 44%(879/2000)๊ฐ€ teacher rollout์—์„œ ํ•œ ๋ฒˆ๋„ ๋‹ฟ์ง€ ์•Š๋Š”๋‹ค(ํ•„์ž ๊ณ„์‚ฐ).
  • \lvert\mathcal{E}_i\rvert๊ฐ€ ๊ฐ™์œผ๋ฉด index ์ˆœ์œผ๋กœ ๊ณ ๋ฅด๋Š”๋ฐ, B=50์—์„œ๋Š” ์ด ์ˆœ์„œ๊ฐ€ 50๊ฐœ ์ค‘ 23๊ฐœ๋ฅผ ์ •ํ•œ๋‹ค.
  • Per-link force์˜ ํฌ๊ธฐ๋Š” 10^{-4} N๊ณผ 10 N ์‚ฌ์ด๋กœ ์ž๋ฅด๊ณ  ํ‘œ์ค€ํŽธ์ฐจ 0.08 N์˜ Gaussian noise๋ฅผ ๋”ํ•œ๋‹ค. ๊ฒฐ๊ณผ ๊ด€์ธก์€ \mathbb{R}^{18}์ด๋‹ค.

๋ฐฉ๋ฒ• ์ƒ์„ธ 3: tensor layout๊ณผ encoder (ยง3.5, ๋ถ€๋ก .4)


Fig. 3. ์„ ํƒ๋œ taxel์˜ tensor layout. ์†๋ฐ”๋‹ฅ์— ์™ธ๋ ฅ F_ext๊ฐ€ ๊ฐ€ํ•ด์ง„ ์ ‘์ด‰์ด ๊ฐ layout์—์„œ tensor์˜ ์–ด๋А ์นธ์— ๋“ค์–ด๊ฐ€๋Š”์ง€ ๋ณด์—ฌ ์ค€๋‹ค. Random์€ ์ด์›ƒ ๊ด€๊ณ„๊ฐ€ ํฉ์–ด์ง€๊ณ , Band๋Š” ๋งํฌ๋ฅผ ๋”ฐ๋ผ ๋ฑ€ ๋ชจ์–‘ ์ˆœ์„œ๋กœ ์ฑ„์›Œ ์ผ๋ถ€ ์ด์›ƒ์ด ๋ณด์กด๋˜๋ฉฐ, Grid๋Š” ํ‘œ๋ฉด ์œ„์น˜์—์„œ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์นธ์— ๋„ฃ์–ด ์ ‘์ด‰ ๋ชจ์–‘์ด ์œ ์ง€๋œ๋‹ค.

CNN์€ ์ธ์ ‘ํ•œ ์ž…๋ ฅ ์นธ์ด ์ด์›ƒํ•œ ์ธก์ •๊ฐ’์ด๋ผ๊ณ  ๊ฐ€์ •ํ•œ๋‹ค. ์ด๋ฏธ์ง€ pixel์„ ์„ž์œผ๋ฉด CNN์€ ๋ง๊ฐ€์ง€๊ณ  ViT๋Š” ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š๋Š”๋‹ค(Jelassi et al., 2022). taxel์€ ์ •ํ•ด์ง„ ๊ฒฉ์ž๊ฐ€ ์—†๋Š”๋ฐ๋„ ๊ธฐ์กด ์ด‰๊ฐ policy๋Š” ๊ฒฉ์ž ์ž…๋ ฅ์— CNN์„ ์จ ์™”๋‹ค. ๊ทธ๋ž˜์„œ layout์˜ ํšจ๊ณผ๋Š” encoder์— ๋”ฐ๋ผ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ๋‹ค. ์ €์ž๋“ค์€ placement๊ฐ€ ๋ผ์–ด๋“ค์ง€ ์•Š๋„๋ก 2000๊ฐœ ์ „๋ถ€๋ฅผ ์“ด ์ƒํƒœ์—์„œ layout๊ณผ encoder๋งŒ ๋ฐ”๊พผ๋‹ค. ๊ฐ ๋งํฌ๋Š” ๊ณ ์ •๋œ ์นธ ๋ธ”๋ก์„ ์ฐจ์ง€ํ•˜๊ณ , ์„ธ layout์€ ์–ด๋А taxel์„ ์–ด๋А ์นธ์— ๋„ฃ๋А๋ƒ๋งŒ ๋‹ค๋ฅด๋‹ค.

encoder spatial module (๋งํฌ ๋ธ”๋ก๋ณ„) temporal taxel์„ ๋‹ค๋ฃจ๋Š” ๋ฐฉ์‹
LSTM ์—†์Œ (๋ชจ๋“  ๋ธ”๋ก์„ ํ•œ ๋ฒกํ„ฐ๋กœ flatten) LSTM(64) ์นธ๋งˆ๋‹ค ๊ณ ์œ  ์ž…๋ ฅ ๊ฐ€์ค‘์น˜, ์ธ์ ‘์„ฑ ๊ฐ€์ • ์—†์Œ
CNN Conv3ร—3(16)โ€“LNโ€“Conv3ร—3(32)โ€“LN, ๋ธ”๋ก๋ณ„ flatten โ†’ ๋งํฌ๋ณ„ linear โ†’ 32-D GRU(64) ์นธ์˜ ์œ„์น˜์™€ ์ธ์ ‘์„ฑ์— ์˜์กด
Link-wise MLP FC(64)โ€“FC(64) ๊ณต์œ , ELU, ๋ธ”๋ก ์•ˆ ํ‰๊ท  โ†’ ๋งํฌ ๊ฐ„ ํ‰๊ท  GRU(64) ์นธ ๋ฐฐ์ •์— ๋ถˆ๋ณ€, taxel ์ •์ฒด์„ฑ์„ ์ง€์›€
๊ณต์œ  ๋ชจ๋“ˆ proprioceptive encoder LSTM(128)โ€“FC(64), policy head FC(512)โ€“FC(256)โ€“FC(128)

๊ฐ ์นธ์€ x_{i,t}์˜ 6์ฑ„๋„์„ ๋‹ด๊ณ , ์„ธ tactile encoder๋Š” ๋ชจ๋‘ FC(32)๋กœ ๋๋‚˜ 64-D proprioceptive feature์™€ ํ•จ๊ป˜ policy head๋กœ ๋“ค์–ด๊ฐ„๋‹ค(Table 9). Link-wise MLP๋Š” ํ‰๊ท ์ด ์นธ ๋ฐฐ์ •์— ๋ถˆ๋ณ€์ด๋ฏ€๋กœ ์„ธ layout์ด ์ด encoder์—๊ฒŒ๋Š” ๋˜‘๊ฐ™์€ ์ž…๋ ฅ์ด๋‹ค. ์ €์ž๋“ค์€ Table 3์—์„œ ์ด encoder์˜ ์„ธ ์นธ ์‚ฌ์ด ์ฐจ์ด๊ฐ€ seed ๋ถ„์‚ฐ์„ ๋ฐ˜์˜ํ•œ๋‹ค๊ณ  ์ง์ ‘ ๋ฐํžŒ๋‹ค. ์ด ๋ฌธ์žฅ์€ ๋’ค์—์„œ ์ด ๋…ผ๋ฌธ์˜ ์žก์Œ ์ˆ˜์ค€์„ ๊ฐ€๋Š ํ•˜๋Š” ์ค‘์š”ํ•œ ๊ธฐ์ค€์ ์ด ๋œ๋‹ค.

์ง๊ด€: 0์€ ์ •๋ณด๊ฐ€ ์•„๋‹ˆ๋‹ค

์ด ๋…ผ๋ฌธ์˜ ์„ค๋ช…์€ ๋‹จ์ˆœํ•˜๋‹ค. ์ด‰๊ฐ ์‹ ํ˜ธ๋Š” ์ ‘์ด‰์ด ์žˆ๋Š” ๊ณณ์—์„œ๋งŒ ์ •๋ณด๋ฅผ ๋‚ธ๋‹ค. ํ›„๋ณด์˜ 44%๊ฐ€ teacher rollout์—์„œ ํ•œ ๋ฒˆ๋„ ๋‹ฟ์ง€ ์•Š๋Š” ์†์ด๋ผ๋ฉด, ํ‘œ๋ฉด์ ์— ๋น„๋ก€ํ•ด ๋ฝ‘์€ taxel์˜ ์ƒ๋‹น์ˆ˜๋Š” ์ฒ˜์Œ๋ถ€ํ„ฐ 0๋งŒ ๋‚ผ ์šด๋ช…์ด๋‹ค. uniform์œผ๋กœ B๋ฅผ ๋Š˜๋ฆฌ๋Š” ๊ฒƒ์€ โ€œ์ •๋ณด๋ฅผ ๋Š˜๋ฆฌ๋Š” ๊ฒƒโ€์ด ์•„๋‹ˆ๋ผ โ€œ0์ด ์•„๋‹Œ taxel์ด ์„ž์—ฌ ๋“ค์–ด์˜ฌ ํ™•๋ฅ ์„ ๋Š˜๋ฆฌ๋Š” ๊ฒƒโ€์— ๊ฐ€๊น๋‹ค. Table 10์˜ Contact Events ์—ด์ด ์ด๋ฅผ ๊ทธ๋Œ€๋กœ ๋ณด์—ฌ ์ค€๋‹ค. ๊ณผ์ œ ์ •๋ณด๋ฅผ ์“ฐ๋Š” ๊ทœ์น™์€ B=10์—์„œ ์ด๋ฏธ ์ ‘์ด‰ ์ด๋ฒคํŠธ์˜ 84~89%๋ฅผ ์žก๋Š”๋ฐ, Uniform์€ B=10์—์„œ 0.1%, B=100์—์„œ๋„ 72.3%๋‹ค.

๋ฐ˜๋Œ€ ๋ฐฉํ–ฅ์˜ ์ง๊ด€๋„ ์žˆ๋‹ค. uniform์ด B๋ฅผ ๋Š˜๋ฆฌ๋ฉด ๊ฒฐ๊ตญ ๋”ฐ๋ผ์žก๋Š” ์ด์œ ๋Š” ์ƒˆ๋กœ ๋“ค์–ด์˜จ taxel์ด ์„œ๋กœ ๊ฐ€๊นŒ์›Œ์ง€๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. Uniform์˜ redundancy(๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์„ ํƒ ์ด์›ƒ๋„ ๊ฐ™์€ ์ ‘์ด‰์„ ์žก๋Š” ๋น„์œจ)๋Š” B=25์—์„œ 8.4%, 100์—์„œ 25.2%, 300์—์„œ 35.0%๋กœ ์˜ค๋ฅธ๋‹ค. ๋งŽ์ด ๊น”๋ฉด ๊ฒฐ๊ตญ ์ž์ฃผ ๋‹ฟ๋Š” ๊ณณ์—๋„ taxel์ด ๋“ค์–ด๊ฐ€๊ณ , ๊ทธ ๋’ค๋กœ๋Š” ์ถ”๊ฐ€ taxel์˜ ์ •๋ณด ์ด๋“์ด ์ค„์–ด๋“ ๋‹ค. ์ด๊ฒƒ์ด Tactile Genesis์˜ โ€œcoverage๊ฐ€ resolution๋ณด๋‹ค ์ค‘์š”ํ•˜๋‹คโ€์™€ ๊ฐ™์€ ๋ฐฉํ–ฅ์ด๋ผ๋Š” ๊ฒƒ์ด ์ €์ž๋“ค์˜ ํ•ด์„์ด๋‹ค.

ํ‘œํ˜„ ์ชฝ์˜ ์ง๊ด€์€ โ€œ์œ„์น˜๊ฐ€ ์•„๋‹ˆ๋ผ ์ •์ฒด์„ฑโ€์ด๋‹ค. LSTM์€ ์นธ๋งˆ๋‹ค ๊ณ ์œ ํ•œ ์ž…๋ ฅ ๊ฐ€์ค‘์น˜๋กœ ์ฝ์œผ๋ฏ€๋กœ ๊ณ ์ •๋œ permutation์€ ํ•™์Šต์œผ๋กœ ํก์ˆ˜๋œ๋‹ค. ๊ฐ™์€ ์ ‘์ด‰์ด ๋Š˜ ๊ฐ™์€ ์นธ์— ๋“ค์–ด์˜ค๊ธฐ๋งŒ ํ•˜๋ฉด ๊ทธ ์นธ์ด tensor์˜ ์–ด๋””์— ์žˆ๋“  ์ƒ๊ด€์—†๋‹ค. ๋ฐ˜๋ฉด Link-wise MLP๋Š” ํ‰๊ท ์„ ๋‚ด๋ฉด์„œ โ€œ์–ด๋А taxel์ด ์ ‘์ด‰์„ ๋ณด๊ณ ํ–ˆ๋Š”๊ฐ€โ€๋ฅผ ์ง€์šด๋‹ค.

์‹คํ—˜: ํ‰๊ฐ€ ํ”„๋กœํ† ์ฝœ (ยง4.1, ๋ถ€๋ก .6)

  • ๊ฐ policy๋Š” 1024 ์—ํ”ผ์†Œ๋“œ๋กœ ํ‰๊ฐ€ํ•˜๊ณ , ๊ฒฐ๊ณผ๋Š” seed 3๊ฐœ์˜ ํ‰๊ท  ยฑ ํ‘œ์ค€ํŽธ์ฐจ๋‹ค.
  • ํ•™์Šต์„ ์–ด๋””์„œ ๋ฉˆ์ถ”๋А๋ƒ์— ๋”ฐ๋ผ ๊ฒฐ๊ณผ๊ฐ€ ๋‹ฌ๋ผ์งˆ ์ˆ˜ ์žˆ์œผ๋ฏ€๋กœ(Mandlekar et al., 2021) ๋ชจ๋“  ํ•™์ƒ์„ ๊ฐ™์€ ํ•™์Šต iteration์—์„œ action ๋ถ„ํฌ์˜ ํ‰๊ท ์œผ๋กœ ํ‰๊ฐ€ํ•œ๋‹ค. ํ™˜๊ฒฝ๋งˆ๋‹ค ์ฒซ ์—ํ”ผ์†Œ๋“œ๋งŒ ์„ผ๋‹ค. ์—ํ”ผ์†Œ๋“œ ๊ธธ์ด์™€ randomization์€ ํ•™์Šต๊ณผ ๊ฐ™๊ณ , teacher๋„ ๊ฐ™์€ ํ”„๋กœํ† ์ฝœ๋กœ ํ‰๊ฐ€ํ•œ๋‹ค.
  • In-Palm Rotation ์„ฑ๊ณต: ํšŒ์ „์ถ• ์˜ค์ฐจ 0.1 rad ๋ฏธ๋งŒ, ์ถ• ์™ธ ์˜ค์ฐจ \pi/6 ๋ฏธ๋งŒ์ด๋ฉด ๋ชฉํ‘œ ๋„๋‹ฌ์ด๊ณ  ๋‹ค์Œ 90๋„ ๋ชฉํ‘œ๊ฐ€ ์„ค์ •๋œ๋‹ค. ๋‚™ํ•˜ ์ „์— ๋ชฉํ‘œ๋ฅผ ํ•˜๋‚˜๋ผ๋„ ๋„๋‹ฌํ•˜๋ฉด ์„ฑ๊ณต์ด๋‹ค.
  • Grasp-and-Lift held fraction: ๋ฌผ์ฒด๊ฐ€ ๊ณ ์ • ๋ฌธํ„ฑ ์œ„๋กœ ๋“ค๋ฆฐ ๋’ค๋ถ€ํ„ฐ, ๋ฌธํ„ฑ ์œ„์— ๋จธ๋ฌธ timestep์˜ ๋น„์œจ์ด๋‹ค. ํ•œ ๋ฒˆ๋„ ๋“ค์ง€ ๋ชปํ•œ ์—ํ”ผ์†Œ๋“œ๋Š” ์ œ์™ธํ•œ๋‹ค.
  • ์™ธ๋ž€: ๋ฌผ์ฒด ์งˆ๋Ÿ‰์ค‘์‹ฌ์— ๊ฐ€ํ•˜๋Š” wrench๋กœ, ํž˜ยท๋ ˆ๋ฒ„ ์˜คํ”„์…‹์ด ์žˆ๋Š” ํž˜ยท์ˆœ์ˆ˜ ํ† ํฌ ์ค‘ ํ•˜๋‚˜๋ฅผ ๊ฐ™์€ ํ™•๋ฅ ๋กœ, ๋ฌด์ž‘์œ„ ๋‹จ์œ„ ๋ฐฉํ–ฅ์œผ๋กœ ๊ณ ๋ฅธ๋‹ค. ๊ฐ•๋„ ๊ณ„์ˆ˜ 0.25๋Š” ๋ฌผ์ฒด ์งˆ๋Ÿ‰๊ณผ ๋ฌด๊ด€ํ•˜๊ฒŒ ํž˜ 1~2 N, ํ† ํฌ 0.02~0.035 Nยทm์— ํ•ด๋‹นํ•œ๋‹ค. ๊ฐ ํŽ„์Šค๋Š” 0.1~0.5 s์˜ half-sine์ด๊ณ  ๋‹ค์Œ ํŽ„์Šค๋Š” ๋๋‚œ ๋’ค 2~4 s์— ์‹œ์ž‘ํ•œ๋‹ค. ์ฒซ ํŽ„์Šค๋Š” 1 s ๋’ค, ๋ฌผ์ฒด๊ฐ€ reset ๋†’์ด๋ณด๋‹ค 5 cm ๋“ค๋ฆฐ ๋‹ค์Œ์— ๊ฐ€ํ•œ๋‹ค. ํ•™์Šต์—์„œ๋Š” ์—ํ”ผ์†Œ๋“œ์˜ 25%์— ์™ธ๋ž€์„ ์ฃผ์ง€ ์•Š์ง€๋งŒ ํ‰๊ฐ€์—์„œ๋Š” ๋ชจ๋“  ์—ํ”ผ์†Œ๋“œ์— ์ค€๋‹ค.

์‹คํ—˜ 1: In-Palm Rotation์˜ placement (ยง4.2, Table 1)

Budget B Uniform Top-contact Contact-cover Link-prop.
No tactile: 0.0 ยฑ 0.0
Least-contact (B=50ยท100): 0.0 ยฑ 0.0
Per-link force: 4.4 ยฑ 1.0
10 0.0 ยฑ 0.0 41.4 ยฑ 0.6 38.7 ยฑ 2.4 0.0 ยฑ 0.0
15 0.0 ยฑ 0.0 42.5 ยฑ 7.6 42.3 ยฑ 1.0 0.0 ยฑ 0.0
25 12.4 ยฑ 1.3 50.5 ยฑ 1.3 50.0 ยฑ 1.8 8.6 ยฑ 3.1
50 27.5 ยฑ 1.5 53.3 ยฑ 3.7 52.8 ยฑ 5.6 40.8 ยฑ 4.3
75 35.5 ยฑ 6.7 59.5 ยฑ 1.2 54.8 ยฑ 1.8 47.9 ยฑ 3.5
100 51.4 ยฑ 2.3 56.6 ยฑ 2.9 55.7 ยฑ 3.8 48.8 ยฑ 0.9
300 56.2 ยฑ 3.6 58.0 ยฑ 2.8 59.4 ยฑ 2.8 56.9 ยฑ 1.3
All (2000): 61.5 ยฑ 0.3
Teacher \pi_T: 93.8

Table 1. ์„ฑ๊ณต๋ฅ (%). XHand1 In-Palm Rotation, LSTM encoder, 3 seeds ํ‰๊ท  ยฑ ํ‘œ์ค€ํŽธ์ฐจ. ์ฐธ์กฐ ํ–‰(๊ธฐ์šธ์ž„)์€ ์› ํ‘œ์—์„œ ๋„ค ์—ด์„ ํ•ฉ์นœ ํ•œ ์นธ์ด๋ผ ์ฒซ ์—ด์— ๊ฐ’์„ ํ•จ๊ป˜ ์ ์—ˆ๋‹ค. ๊ตต๊ฒŒ๋Š” budget๋ณ„ ์ตœ๊ณ , ๊ธฐ์šธ์ž„์€ 2์œ„(์› ํ‘œ์˜ ์ƒ‰ ๊ฐ•์กฐ๋ฅผ ์˜ฎ๊น€).

๋ณธ๋ฌธ์ด ์ด ํ‘œ์—์„œ ๋Œ์–ด๋‚ด๋Š” ์ฃผ์žฅ์€ ๋„ค ๊ฐ€์ง€๋‹ค.

  1. โ€œ๋ช‡ ๊ฐœ๋ณด๋‹ค ์–ด๋А ๊ฒƒโ€: Top-contactยทContact-cover๋Š” teacher๊ฐ€ ๋‹ฟ๋Š” ๊ณณ์— taxel์„ ๋‘๋ฏ€๋กœ ์ ์€ ์ˆ˜๋กœ ์ ‘์ด‰ ์ด๋ฒคํŠธ ๋Œ€๋ถ€๋ถ„์„ ์žก๋Š”๋‹ค. ํ‘œ๋กœ ํ™•์ธํ•˜๋ฉด Top-contact B=25(50.5 ยฑ 1.3)์™€ Uniform B=100(51.4 ยฑ 2.3)์ด ๊ฐ™๊ณ , Contact-cover B=10(38.7)์ด Uniform B=75(35.5)๋ณด๋‹ค ๋†’์œผ๋ฉฐ, Top-contact B=75(59.5)๋Š” Uniform B=300(56.2)์— ๋’ค์ง€์ง€ ์•Š๋Š”๋‹ค(์ฐจ์ด๋Š” 3 seeds ๋ถ„์‚ฐ ์•ˆ์ชฝ). ๊ฐ™์€ ์„ฑ๊ณต๋ฅ ์„ ์–ป๋Š” ๋ฐ uniform์ด 4๋ฐฐ ์ด์ƒ์˜ taxel์„ ์“ด๋‹ค.
  2. ๋“œ๋ฌผ๊ฒŒ ๋‹ฟ๋Š” taxel = ์ด‰๊ฐ ์—†์Œ: Least-contact๋Š” ํ•œ ๋ฒˆ๋„ ์„ฑ๊ณตํ•˜์ง€ ๋ชปํ•ด ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ๊ณผ ๊ฐ™๋‹ค. Uniform๋„ ๊ฐ€์žฅ ๋‚ฎ์€ budget์—์„œ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ ์‹คํŒจํ•˜๋Š”๋ฐ, ๊ทธ๋•Œ ๊ณ ๋ฅธ taxel์€ teacher ์ ‘์ด‰์„ ๊ฑฐ์˜ ์žก์ง€ ๋ชปํ•œ๋‹ค(Table 10์˜ Contact Events 0.1~0.2%).
  3. ์ด‰๊ฐ์€ ํšŒ์ „์„ ์ด๋ˆ๋‹ค: ์ด‰๊ฐ์ด ์—†์œผ๋ฉด ํ•™์ƒ์€ ๋Œ€๋ถ€๋ถ„ ๋ฌผ์ฒด๋ฅผ ์ฅ๊ณ  ์žˆ์ง€๋งŒ ๋Œ๋ฆฌ์ง€ ๋ชปํ•œ๋‹ค. ๊ณต๊ฐ„ ํ•ด์ƒ๋„ ์—†๋Š” Per-link force๋Š” ๊ฑฐ์˜ ๊ฐœ์„ ์ด ์—†์œผ๋ฏ€๋กœ(4.4%) ๊ณต๊ฐ„ ๋ถ„ํ•ด๋œ ์ ‘์ด‰ ์ •๋ณด๊ฐ€ ํ•„์š”ํ•˜๋‹ค.
  4. ํฌํ™”์™€ ์ˆ˜๋ ด: ๊ณผ์ œ ์ •๋ณด๋ฅผ ์“ฐ๋Š” ๊ทœ์น™์€ ์ค‘๊ฐ„ budget๋ถ€ํ„ฐ ํฌํ™”ํ•˜๊ณ , Uniform์€ B=300์—์„œ ๋”ฐ๋ผ์žก์•„ ๋ชจ๋“  ๋ฐฉ๋ฒ•์ด 2000๊ฐœ ์ „๋ถ€(61.5)์— ๊ฐ€๊นŒ์šด ๊ฐ’์œผ๋กœ ๋ชจ์ธ๋‹ค. Link-proportional์€ ๋“œ๋ฌผ๊ฒŒ ๋‹ฟ๋Š” ๋งํฌ์˜ taxel์„ ์—„์ง€ ๋์œผ๋กœ ์˜ฎ๊ธฐ๋Š”๋ฐ(Fig. 2), ์ค‘๊ฐ„ budget์—์„œ๋Š” Uniform๋ณด๋‹ค ๋‚ซ๊ณ  B\ge100์—์„œ๋Š” ๊ฐ™์œผ๋ฉฐ, B\le25์—์„œ๋Š” Uniform์ฒ˜๋Ÿผ ๋‚˜์˜๋‹ค. 2000๊ฐœ๋ฅผ ๋‹ค ์จ๋„ teacher(93.8)์™€์˜ ๊ฒฉ์ฐจ๊ฐ€ ๋‚จ๋Š”๋‹ค.

๋ถ€๋ก .7.1: ์ ‘์ด‰ coverage (Table 10)

์ง€ํ‘œ B Uniform Top-contact Contact-cover Link-prop.
Contact Events (%) 10 0.1 83.9 88.8 0.1
15 0.2 84.7 94.8 0.1
25 24.6 85.6 98.1 25.1
50 35.9 87.5 99.7 54.4
75 45.8 88.9 99.9 59.5
100 72.3 93.7 99.9 64.4
300 91.8 99.8 100.0 84.5
Taxel Readings (%) 10 0.0 8.1 5.1 0.0
15 0.1 14.5 7.6 0.1
25 0.9 26.4 12.7 0.9
50 1.7 48.7 18.3 2.7
75 3.4 63.1 26.9 4.5
100 5.5 72.8 31.7 7.9
300 14.4 98.0 52.6 17.4
Redundancy (%) 25 8.4 95.7 76.1 11.1
100 25.2 91.3 75.9 29.7
300 35.0 79.7 76.6 44.0

Table 10. XHand1 In-Palm Rotation. Contact Events๋Š” ์ ‘์ด‰ ์ค‘์ธ ๋งํฌ ์œ„์˜ ์„ ํƒ๋œ taxel ์ค‘ ํ•˜๋‚˜๋ผ๋„ 0์ด ์•„๋‹Œ ๊ฐ’์„ ๋‚ธ ์ ‘์ด‰ ์ด๋ฒคํŠธ์˜ ๋น„์œจ, Taxel Readings๋Š” teacher rollout์˜ 0์ด ์•„๋‹Œ ํŒ๋… ์ „์ฒด ์ค‘ ์„ ํƒ๋œ taxel์— ๋–จ์–ด์ง„ ๋น„์œจ, Redundancy๋Š” ํ•œ taxel์˜ ์ ‘์ด‰์„ ๊ฐ€์žฅ ๊ฐ€๊นŒ์šด ์„ ํƒ ์ด์›ƒ๋„ ์žก๋Š” ๋น„์œจ์ด๋‹ค.

์„ธ ์ง€ํ‘œ๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ๊ฒƒ์„ ์žฐ๋‹ค๋Š” ์ ์ด ์ด ํ‘œ๋ฅผ ์ฝ๋Š” ์š”๋ น์ด๋‹ค. Contact-cover๋Š” B=300์—์„œ ์ ‘์ด‰ ์ด๋ฒคํŠธ๋ฅผ 100% ์žก์ง€๋งŒ ํŒ๋…์€ ์ ˆ๋ฐ˜ ์ •๋„(52.6%)๋งŒ ์žก๋Š”๋‹ค. ์™„์ „ coverage ์ดํ›„์— ์ถ”๊ฐ€๋˜๋Š” taxel์ด index ์ˆœ์„œ๋ฅผ ๋”ฐ๋ฅด๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. Top-contact๋Š” ๊ฐ€์žฅ ์ž‘์€ budget๋ถ€ํ„ฐ redundancy๊ฐ€ ๋†’์ง€๋งŒ(95.7%) ๊ทธ ์ค‘๋ณต์ด ๊ฐ€์žฅ ์ž์ฃผ ๋‹ฟ๋Š” ์ง€์ ์— ์žˆ์œผ๋ฏ€๋กœ ํŒ๋… ๋น„์œจ์€ ๋ชจ๋“  budget์—์„œ ๊ฐ€์žฅ ๋†’๋‹ค. Uniform์€ budget์ด ์ปค์งˆ์ˆ˜๋ก redundancy๊ฐ€ ์˜ค๋ฅธ๋‹ค. ์„ฑ๊ณต๋ฅ (Table 1)๊ณผ ๋‚˜๋ž€ํžˆ ๋†“์œผ๋ฉด ์„ฑ๊ณต๋ฅ ์„ ๊ฐ€์žฅ ์ž˜ ๋”ฐ๋ผ๊ฐ€๋Š” ๊ฒƒ์€ Contact Events๋‹ค. UniformยทLink-prop. B=10ยท15์˜ 0.1~0.2%๋Š” ์„ฑ๊ณต๋ฅ  0.0%์™€, B=25์˜ ์•ฝ 25%๋Š” ์„ฑ๊ณต๋ฅ  8.6~12.4%์™€ ๋งž๋ฌผ๋ฆฐ๋‹ค.

๋ถ€๋ก .7.2: ์—ํ”ผ์†Œ๋“œ ๊ฒฐ๊ณผ (Table 11)

B Placement Success Drop Under-rotation
0 No tactile 0.0 ยฑ 0.0 15.0 ยฑ 12.7 85.0 ยฑ 12.7
25 Uniform 12.4 ยฑ 1.3 30.8 ยฑ 5.1 56.8 ยฑ 4.5
Top-contact 50.5 ยฑ 1.3 16.2 ยฑ 1.1 33.3 ยฑ 3.2
Contact-cover 50.0 ยฑ 1.8 14.6 ยฑ 3.6 35.4 ยฑ 4.5
Link-prop. 8.6 ยฑ 3.1 30.2 ยฑ 5.3 61.2 ยฑ 4.7
100 Uniform 51.4 ยฑ 2.3 16.5 ยฑ 3.2 32.1 ยฑ 1.1
Top-contact 56.6 ยฑ 2.9 14.1 ยฑ 1.0 29.3 ยฑ 0.3
Contact-cover 55.7 ยฑ 3.8 16.8 ยฑ 1.5 27.5 ยฑ 4.3
Link-prop. 48.8 ยฑ 0.9 18.2 ยฑ 1.7 33.0 ยฑ 3.3
300 Uniform 56.2 ยฑ 3.6 13.9 ยฑ 2.0 29.9 ยฑ 0.1
Top-contact 58.0 ยฑ 2.8 13.8 ยฑ 0.8 28.2 ยฑ 3.2
Contact-cover 59.4 ยฑ 2.8 15.3 ยฑ 0.6 25.3 ยฑ 1.9
Link-prop. 56.9 ยฑ 1.3 15.4 ยฑ 1.2 27.7 ยฑ 3.9
2000 All 61.5 ยฑ 0.3 13.1 ยฑ 1.3 25.4 ยฑ 2.8

Table 11. XHand1 In-Palm Rotation, LSTM encoder(%). Drop์€ ์ฒซ ๋ชฉํ‘œ ์ „์— ๋–จ์–ด์ง„ ๊ฒฝ์šฐ, Under-rotation์€ ๋ชฉํ‘œ ๋„๋‹ฌ๋„ ๋‚™ํ•˜๋„ ์—†์ด ๋๋‚œ ๊ฒฝ์šฐ๋‹ค. ๊ฐ ํ–‰์˜ ์„ธ ๊ฐ’์˜ ํ•ฉ์€ 100.0์ด๊ณ  Success ์—ด์€ Table 1๊ณผ ์ผ์น˜ํ•œ๋‹ค(ํ•„์ž ํ™•์ธ).

๋‚™ํ•˜์œจ์€ ์ด‰๊ฐ ์œ ๋ฌด์™€ ์ƒ๊ด€์—†์ด ๋น„์Šทํ•˜๋‹ค(13.1~18.2%). ์˜ˆ์™ธ๋Š” B=25์˜ Uniform(30.8%)๊ณผ Link-prop.(30.2%)๋ฟ์ธ๋ฐ, ์ด ๋‘˜์€ ๊ทธ budget์—์„œ ์ ‘์ด‰ ์ด๋ฒคํŠธ๋ฅผ 25% ์ •๋„๋งŒ ์žก๋Š”๋‹ค. ์ €์ž๋“ค์€ ์ด๋ฅผ โ€œ๋ถ€์กฑํ•œ ์ ‘์ด‰ ์ •๋ณด๋กœ ํšŒ์ „์„ ์‹œ๋„ํ•˜๋‹ค ๋–จ์–ด๋œจ๋ฆฐ๋‹คโ€๋กœ ์ฝ๋Š”๋‹ค. ์ด‰๊ฐ์€ under-rotation์„ ์ค„์ด๊ณ (85.0 โ†’ 25.4) ๊ทธ ๊ฐ์†Œ๋Š” budget๊ณผ ํ•จ๊ป˜ ์ปค์ง„๋‹ค. ๊ทธ๋ž˜์„œ ์ด ๊ณผ์ œ์—์„œ ์ด‰๊ฐ์€ ๋ฌผ์ฒด๋ฅผ ์ฅ๋Š” ๊ฒƒ๋ณด๋‹ค ํšŒ์ „์„ ์ง„ํ–‰์‹œํ‚ค๋Š” ๋ฐ ์“ฐ์ธ๋‹ค๋Š” ๊ฒƒ์ด ์ €์ž๋“ค์˜ ๊ฒฐ๋ก ์ด๋‹ค. ๋ง๋ถ™์ด๋ฉด ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ์˜ ๋‚™ํ•˜์œจ 15.0 ยฑ 12.7์€ seed ๊ฐ„ ๋ถ„์‚ฐ์ด ๋งค์šฐ ์ปค์„œ, โ€œB=25 Uniform์ด ์ด‰๊ฐ ์—†์„ ๋•Œ๋ณด๋‹ค ๋” ์ž์ฃผ ๋–จ์–ด๋œจ๋ฆฐ๋‹คโ€๋Š” ๋น„๊ต์˜ ๊ธฐ์ค€์  ์ž์ฒด๊ฐ€ ํ”๋“ค๋ฆฐ๋‹ค. ๋ณธ๋ฌธ์€ โ€œ๊ฐ€์žฅ ๋‚ฎ์€ budget์—์„œโ€๋ผ๊ณ  ์“ฐ์ง€๋งŒ Table 11์—๋Š” B=10ยท15 ํ–‰์ด ์—†๋‹ค.

์‹คํ—˜ 2: Grasp-and-Lift์—์„œ์˜ placement (ยง4.2, Table 2)

Budget B Uniform Top-contact Contact-cover Link-prop.
Sharpa ยท No tactile: 9.4 ยฑ 0.7
25 15.9 ยฑ 0.7 22.3 ยฑ 6.7 27.9 ยฑ 8.5 21.2 ยฑ 3.9
50 24.0 ยฑ 0.8 29.1 ยฑ 12.5 38.9 ยฑ 15.7 26.6 ยฑ 5.5
100 25.6 ยฑ 3.4 24.3 ยฑ 1.3 32.0 ยฑ 5.4 26.9 ยฑ 2.8
300 42.6 ยฑ 2.1 27.0 ยฑ 3.7 57.0 ยฑ 4.7 59.8 ยฑ 2.2
Teacher \pi_T: 96.0
Allegro ยท No tactile: 22.9 ยฑ 2.1
25 39.5 ยฑ 4.6 22.8 ยฑ 3.5 45.5 ยฑ 3.1 40.0 ยฑ 9.2
50 43.3 ยฑ 3.4 38.2 ยฑ 1.0 45.9 ยฑ 8.8 46.1 ยฑ 3.1
100 59.0 ยฑ 16.1 60.4 ยฑ 3.2 61.3 ยฑ 0.7 52.9 ยฑ 7.1
300 57.0 ยฑ 11.9 57.2 ยฑ 1.6 72.2 ยฑ 2.1 54.2 ยฑ 4.2
Teacher \pi_T: 93.0

Table 2. Held fraction(%). Grasp-and-Lift, LSTM encoder. ์ฐธ์กฐ ํ–‰(๊ธฐ์šธ์ž„)์€ ์ฒซ ์—ด์— ๊ฐ’์„ ์ ์—ˆ๋‹ค. ๊ตต๊ฒŒ๋Š” ์†ยทbudget๋ณ„ ์ตœ๊ณ , ๊ธฐ์šธ์ž„์€ 2์œ„. ์› ํ‘œ ์บก์…˜์—๋Š” seed ์ˆ˜๊ฐ€ ์—†์ง€๋งŒ ๋ถ€๋ก .6์€ ๋ชจ๋“  ๊ฒฐ๊ณผ๋ฅผ 3 seeds ํ‰๊ท  ยฑ ํ‘œ์ค€ํŽธ์ฐจ๋กœ ๋ณด๊ณ ํ•œ๋‹ค๊ณ  ์ ๋Š”๋‹ค.

์ด ๊ณผ์ œ์—์„œ๋Š” ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ์„ ํฌํ•จํ•œ ๋ชจ๋“  ํ•™์ƒ์ด 1 s ์•ˆ์— ์ฒซ ์ ‘์ด‰์„ ํ•˜๊ณ  ๊ฑฐ์˜ ๋ชจ๋“  ์—ํ”ผ์†Œ๋“œ์—์„œ ๋ฌผ์ฒด๋ฅผ ๋“ ๋‹ค. ์ด‰๊ฐ์˜ ์ด๋“์€ ์™ธ๋ž€ ์†์—์„œ ๋ฒ„ํ‹ฐ๋Š” ๋ฐ์„œ ๋‚˜์˜จ๋‹ค. ๋‘ ์†์€ ์ด‰๊ฐ ์—†์ด๋„ held fraction์ด ๋‹ค๋ฅด๋‹ค(Sharpa 9.4, Allegro 22.9).

placement์˜ ํšจ๊ณผ๋Š” In-Palm Rotation๊ณผ ๋ฐ˜๋Œ€ ๊ฒฝํ–ฅ์ด๋‹ค. Top-contact๋ฅผ ๋นผ๋ฉด ๋‚ฎ์€ budget์—์„œ๋Š” ๋ฐฉ๋ฒ•๋ผ๋ฆฌ ์ฐจ์ด๊ฐ€ ์ž‘๊ณ , ๊ฐ€์žฅ ํฐ budget์—์„œ ๊ฐˆ๋ฆฐ๋‹ค. ๊ทธ budget์—์„œ Contact-cover๋Š” ๋‘ ์† ๋ชจ๋‘ Uniform์„ ๋„˜๊ณ , Top-contact๋Š” Sharpa์—์„œ Uniform๋ณด๋‹ค ํฌ๊ฒŒ ๋‚ฎ๋‹ค(27.0 vs 42.6). ๊ฐ€์žฅ ๋งŽ์ด ๋‹ฟ๋Š” ์ง€์ ์— taxel์„ ๋ชฐ์•„ ์ฃผ๋Š” ๊ฒƒ์€ ์ด๋“์ด ์—†๋‹ค. Top-contact๋Š” Uniform์„ ์ผ๊ด€๋˜๊ฒŒ ์ด๊ธฐ์ง€ ๋ชปํ•˜๊ณ , Allegro B=25์—์„œ๋Š” ์ด‰๊ฐ ์—†๋Š” ํ•™์ƒ๊ณผ ๋‹ค๋ฅด์ง€ ์•Š๋‹ค(22.8 vs 22.9). ์ €์ž๋“ค์€ ์ด ๊ณผ์ œ์—์„œ teacher ์ ‘์ด‰์ด ๋” ๋งŽ์€ ๋งํฌ์— ํผ์ ธ ์žˆ์œผ๋ฏ€๋กœ โ€œํ•œ ๋งํฌ์— ๋ชฐ๊ธฐโ€๋ณด๋‹ค โ€œ์ ‘์ด‰ํ•œ ๋งํฌ๋ฅผ ๋ฎ๊ธฐโ€๊ฐ€ ์ค‘์š”ํ•˜๋‹ค๊ณ  ์„ค๋ช…ํ•œ๋‹ค.

๋ถ€๋ก .7.3: ์ ‘์ด‰ ์ง‘์ค‘๋„ (Table 12)

๊ณผ์ œ Hand Top-100 Share (%) Entropy Links Spanned
In-Palm Rotation XHand1 73.5 0.70 3 / 6
Grasp-and-Lift Allegro 66.0 0.72 16 / 39
Grasp-and-Lift Sharpa 38.3 0.84 5 / 17

Table 12. ๊ฐ ์†์˜ ํ›„๋ณด 2000๊ฐœ์— ๋Œ€ํ•œ teacher rollout. Top-100 Share๋Š” ๊ฐ€์žฅ ๋งŽ์ด ๋‹ฟ์€ 100๊ฐœ taxel์ด ์ฐจ์ง€ํ•˜๋Š” ์ ‘์ด‰ ํŒ๋… ๋น„์œจ, Entropy๋Š” taxel ์œ„ Shannon entropy๋ฅผ log 2000์œผ๋กœ ์ •๊ทœํ™”ํ•œ ๊ฐ’, Links Spanned๋Š” top-100 taxel์ด ์žˆ๋Š” ๋งํฌ ์ˆ˜ / ํ›„๋ณด taxel์ด ์žˆ๋Š” ๋งํฌ ์ˆ˜.

XHand1๊ณผ Allegro๋Š” ์ƒ์œ„ 100๊ฐœ๊ฐ€ ํŒ๋…์˜ ๋Œ€๋ถ€๋ถ„์„ ์ฐจ์ง€ํ•˜๊ณ  Sharpa๋Š” ์ ˆ๋ฐ˜์ด ์•ˆ ๋œ๋‹ค. Allegro๋Š” XHand1๋งŒํผ ์ง‘์ค‘๋ผ ์žˆ์ง€๋งŒ(66.0 vs 73.5, entropy 0.72 vs 0.70) ๊ทธ ์ƒ์œ„ taxel์ด ์†๋ฐ”๋‹ฅ๊ณผ ์—„์ง€๊ฐ€ ์•„๋‹ˆ๋ผ ๋„ค ์†๊ฐ€๋ฝ์— ๊ฑธ์ณ ์žˆ๋‹ค. ์ด ํ‘œ๋Š” โ€œGrasp-and-Lift์˜ ์ ‘์ด‰์ด ๋” ๋งŽ์€ ๋งํฌ์— ํผ์ง„๋‹คโ€๋Š” ๋ณธ๋ฌธ ์„ค๋ช…์˜ ๊ทผ๊ฑฐ์ธ๋ฐ, ์ ˆ๋Œ€ ๋งํฌ ์ˆ˜(16๊ฐœ, 5๊ฐœ vs 3๊ฐœ)๋กœ๋Š” ๋งž์ง€๋งŒ ๋น„์œจ(Allegro 16/39 โ‰ˆ 41%, Sharpa 5/17 โ‰ˆ 29%, XHand1 3/6 = 50%, ํ•„์ž ๊ณ„์‚ฐ)๋กœ๋Š” XHand1์ด ์˜คํžˆ๋ ค ๊ฐ€์žฅ ๋„“๋‹ค. ์ง‘์ค‘๋„ ์ง€ํ‘œ ์„ธ ๊ฐœ๊ฐ€ ์„œ๋กœ ๋‹ค๋ฅธ ์ˆœ์„œ๋ฅผ ๋‚ธ๋‹ค๋Š” ์ ๋„ ํ•จ๊ป˜ ๊ธฐ์–ตํ•ด ๋‘˜ ๋งŒํ•˜๋‹ค.

โ€œuniform์ด taxel ์ˆ˜๋ฅผ ๊ณผ๋Œ€ํ‰๊ฐ€ํ•œ๋‹คโ€๋Š” ์ฃผ์žฅ์˜ ๊ทผ๊ฑฐ ์กฐ๊ฑด

์ดˆ๋ก๊ณผ ๊ฒฐ๋ก ์˜ ํ—ค๋“œ๋ผ์ธ์€ โ€œUniform placement therefore overestimates the number of taxels a task requiresโ€๋‹ค. ์ด ๋ฌธ์žฅ์„ ์ง์ ‘ ๋ฐ›์น˜๋Š” ์ฆ๊ฑฐ๊ฐ€ ์–ด๋А ์กฐ๊ฑด์—์„œ ๋‚˜์™”๋Š”์ง€๋ฅผ ํ•œ ํ‘œ๋กœ ์ •๋ฆฌํ•œ๋‹ค.

๊ทผ๊ฑฐ ์† ๊ณผ์ œ encoder seeds budget ํŒ์ •
Top-contact B=25 โ‰ˆ Uniform B=100 (50.5 vs 51.4), Top-contact B=75 โ‰ฅ Uniform B=300 (59.5 vs 56.2) XHand1 In-Palm Rotation LSTM 3 10~300 ๊นจ๋—ํ•˜๊ฒŒ ์„ฑ๋ฆฝ
Uniform B=10ยท15 = 0.0%, Contact Events 0.1~0.2% XHand1 In-Palm Rotation LSTM 3 10ยท15 ๊นจ๋—ํ•˜๊ฒŒ ์„ฑ๋ฆฝ
Contact-cover B=100 (61.3 ยฑ 0.7) โ‰ฅ Uniform B=300 (57.0 ยฑ 11.9) Allegro Grasp-and-Lift LSTM 3 25~300 ๋ฐฉํ–ฅ์€ ๊ฐ™์ง€๋งŒ Uniform ๋ถ„์‚ฐ์ด ์ปค์„œ ์•ฝํ•จ
Contact-cover B=100 (32.0) < Uniform B=300 (42.6) Sharpa Grasp-and-Lift LSTM 3 25~300 ์„ฑ๋ฆฝํ•˜์ง€ ์•Š์Œ (B=300๋ผ๋ฆฌ๋Š” Contact-cover๊ฐ€ ์šฐ์œ„)
Top-contact๊ฐ€ Uniform์„ ์ผ๊ด€๋˜๊ฒŒ ๋ชป ์ด๊น€ SharpaยทAllegro Grasp-and-Lift LSTM 3 25~300 ๋ฐ˜๋ก€(๊ทœ์น™์— ๋”ฐ๋ผ ๋‹ค๋ฆ„)

์ฆ‰ โ€œ์ž‘์€ ์ˆ˜๋กœ ํฐ uniform ์ง‘ํ•ฉ์„ ๋งž์ถ˜๋‹คโ€๋Š” ๋ฌธ์žฅ์ด 4๋ฐฐ ์ด์ƒ์˜ ์ฐจ์ด๋กœ ์„ฑ๋ฆฝํ•˜๋Š” ๊ณณ์€ XHand1 In-Palm Rotation, LSTM, 3 seeds, 1024 ์—ํ”ผ์†Œ๋“œ, ๊ฐ™์€ ํ•™์Šต iteration์ด๋‹ค. ์ €์ž๋„ ๊ฒฐ๋ก ์—์„œ โ€œThese lessons come from one hand and one task in which contact concentrates on a few linksโ€๋ผ๊ณ  ๋ฒ”์œ„๋ฅผ ๋ฐํžŒ๋‹ค. Grasp-and-Lift๋Š” Contact-cover์— ํ•œํ•ด ๊ฐ™์€ ๋ฐฉํ–ฅ์˜ ๊ทผ๊ฑฐ๋ฅผ ์ผ๋ถ€ ๋ณดํƒœ์ง€๋งŒ(Allegro), Sharpa์—์„œ๋Š” ๊ฐ™์€ ์„ฑ๊ณต๋ฅ ์„ ์–ป๋Š” ๋ฐ Contact-cover๊ฐ€ ํ•„์š”๋กœ ํ•˜๋Š” taxel ์ˆ˜๊ฐ€ uniform๋ณด๋‹ค ์ ๋‹ค๋Š” ์ฆ๊ฑฐ๊ฐ€ Table 2์— ์—†๋‹ค. ์ดˆ๋ก์ด โ€œthree robot handsโ€๋ฅผ ๋‚ด์„ธ์šฐ์ง€๋งŒ ํ—ค๋“œ๋ผ์ธ ์ฃผ์žฅ์˜ ์ •๋Ÿ‰ ๊ทผ๊ฑฐ๋Š” ํ•œ ์†์— ์žˆ๋‹ค.

์‹คํ—˜ 3: ํ‘œํ˜„ (ยง4.3, Table 3)

Encoder Random Band Grid
CNN 61.4 ยฑ 0.6 64.3 ยฑ 1.3 61.3 ยฑ 1.1
LSTM 60.8 ยฑ 3.9 58.0 ยฑ 3.2 61.5 ยฑ 0.3
Link-wise MLP 33.2 ยฑ 8.6 25.2 ยฑ 3.3 18.6 ยฑ 1.1

Table 3. ์„ฑ๊ณต๋ฅ (%). XHand1 In-Palm Rotation, ํ›„๋ณด 2000๊ฐœ ์ „๋ถ€. ์› ํ‘œ์—๋Š” ๊ฐ•์กฐ ํ‘œ์‹œ๊ฐ€ ์—†๋‹ค.

๊ฐ™์€ taxel, ๊ฐ™์€ ์‹ ํ˜ธ์—์„œ encoder์˜ ํšจ๊ณผ๊ฐ€ layout์˜ ํšจ๊ณผ๋ณด๋‹ค ํ›จ์”ฌ ํฌ๋‹ค. CNN์—์„œ๋Š” Grid๊ฐ€ ํ‘œ๋ฉด ์ด์›ƒ์„ ๋ณด์กดํ•˜๊ณ  Random์ด ๊ทธ๊ฒƒ์„ ํŒŒ๊ดดํ•˜๋Š”๋ฐ๋„ ์„ธ layout์ด ๋น„์Šทํ•˜๋‹ค. pixel์„ ์„ž์œผ๋ฉด CNN์ด ๋ง๊ฐ€์ง€๋Š” ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜์™€ ๋‹ค๋ฅธ ๊ฒฐ๊ณผ๋‹ค. LSTM์—์„œ๋„ Random๊ณผ Grid๊ฐ€ ๋น„์Šทํ•˜๋‹ค. ์ €์ž๋“ค์€ โ€œpolicy ์„ฑ๋Šฅ์€ taxel์˜ ์ •์ฒด์„ฑ์— ๋‹ฌ๋ ค ์žˆ๊ณ  tensor ์•ˆ ์ˆœ์„œ์—๋Š” ์˜ํ–ฅ์„ ๋ฐ›์ง€ ์•Š๋Š”๋‹คโ€๊ณ  ๊ฒฐ๋ก  ๋‚ธ๋‹ค. Link-wise MLP๋Š” CNN๊ณผ recurrent ๋ชจ๋“ˆ(GRU)์„ ๊ณต์œ ํ•˜๊ณ  spatial module๋งŒ ๋‹ค๋ฅธ๋ฐ ๋‘˜๋ณด๋‹ค ํ›จ์”ฌ ๋‚ฎ๋‹ค. ๋งํฌ๋ณ„ ํ‰๊ท ์ด taxel ์ •์ฒด์„ฑ์„ ๋ฒ„๋ฆฌ๊ธฐ ๋•Œ๋ฌธ์ด๊ณ , ์ด๋Š” ยง4.2์˜ โ€œ๊ณต๊ฐ„ ๋ถ„ํ•ด๋œ ์ ‘์ด‰ ์ •๋ณด๊ฐ€ ํ•„์š”ํ•˜๋‹คโ€์™€ ์ผ์น˜ํ•œ๋‹ค.

๋‘ ๊ฐ€์ง€ ๊ต์ฐจ ํ™•์ธ์ด ๊ฐ€๋Šฅํ•˜๋‹ค. ์ฒซ์งธ, LSTM Grid(61.5 ยฑ 0.3)๋Š” Table 1์˜ All(2000)ยทTable 11์˜ All๊ณผ ๊ฐ™์€ ๊ฐ’์ด๋‹ค. ๋”ฐ๋ผ์„œ Table 1์˜ placement ์‹คํ—˜์€ Grid layout์—์„œ ์ˆ˜ํ–‰๋๋‹ค๊ณ  ์ฝํžŒ๋‹ค(ํ•„์ž ์ถ”๋ก , ๋ณธ๋ฌธ์— ๋ช…์‹œ๋Š” ์—†๋‹ค). ๋‘˜์งธ, Link-wise MLP์˜ ์„ธ ์นธ์€ ๊ฐ™์€ ์ž…๋ ฅ์ธ๋ฐ ํ‰๊ท ์ด 18.6์—์„œ 33.2๊นŒ์ง€ 14.6%p ๋ฒŒ์–ด์ง„๋‹ค. ์ €์ž๋Š” ์ด๊ฒƒ์ด seed ๋ถ„์‚ฐ์ด๋ผ๊ณ  ์ ๋Š”๋‹ค. ์ด ๊ฐ’์€ ์•„๋ž˜ ๋น„ํŒ์—์„œ ์“ฐ๋Š” ์žก์Œ ์ˆ˜์ค€์˜ ๊ธฐ์ค€์ ์ด๋‹ค.

์‹คํ—˜ 4: ํ•™์Šต ๋น„์šฉ (ยง4.4, Table 4)

Tactile observation Env-steps/s Speed-up VRAM
None 42,810 7.6ร— 1.5 GB
B=25 32,233 5.7ร— 1.6 GB
B=100 29,503 5.3ร— 1.9 GB
B=300 23,817 4.2ร— 2.5 GB
B=2000 8,411 1.5ร— 8.1 GB
Full (ours) 5,610 1.0ร— 12.4 GB

Table 4. In-Palm Rotation, 2048 ํ™˜๊ฒฝ, NVIDIA H100. Speed-up์€ ์ด ๋…ผ๋ฌธ์ด ์“ด full observation ๋Œ€๋น„.

์„ฑ๊ณต๋ฅ ์ด ํฌํ™”ํ•˜๋Š” budget์—์„œ ์„ ํƒ๋œ taxel๋งŒ ์‹œ๋ฎฌ๋ ˆ์ด์…˜ํ•˜๋ฉด full observation๋ณด๋‹ค ์•ฝ 5๋ฐฐ ๋น ๋ฅด๊ณ  ๋ฉ”๋ชจ๋ฆฌ๋„ ์ผ๋ถ€๋งŒ ์“ด๋‹ค. 2000๊ฐœ ์ „๋ถ€๋ฅผ ์“ฐ๋Š” ๋‘ ํ–‰์€ zero-padding๋œ ์นธ๋งŒ ๋‹ค๋ฅด๋ฉฐ, padding๋งŒ์œผ๋กœ throughput์ด 3๋ถ„์˜ 1 ์ค„๊ณ (8,411 โ†’ 5,610, ํ•„์ž ๊ณ„์‚ฐ โˆ’33%) ๋ฉ”๋ชจ๋ฆฌ๋Š” ์ ˆ๋ฐ˜ ๋Š˜์–ด๋‚œ๋‹ค(8.1 โ†’ 12.4 GB, +53%). ๊ทธ๋ฆฌ๊ณ  ์ด ๋…ผ๋ฌธ์˜ ๋ชจ๋“  policy๋Š” full observation์œผ๋กœ ํ•™์Šตํ–ˆ๋‹ค. ์ฆ‰ Table 4์˜ speed-up์€ ์ด ๋…ผ๋ฌธ์ด ์‹ค์ œ๋กœ ์–ป์€ ์ ˆ๊ฐ์ด ์•„๋‹ˆ๋ผ โ€œ์„ ํƒ์„ ํ•˜๋ฉด ์–ป์„ ์ˆ˜ ์žˆ๋Š”โ€ ๋น„์šฉ์ด๋‹ค. ๋˜ ์ด 5๋ฐฐ์—๋Š” padding ์ œ๊ฑฐ ํšจ๊ณผ๊ฐ€ ์„ž์—ฌ ์žˆ๋‹ค. padding ์—†๋Š” 2000๊ฐœ(B=2000) ๋Œ€๋น„๋กœ ๋‹ค์‹œ ๊ณ„์‚ฐํ•˜๋ฉด B=100์€ 3.5๋ฐฐ, B=25๋Š” 3.8๋ฐฐ๋‹ค(ํ•„์ž ๊ณ„์‚ฐ).

๊ฒฐ๋ก ๊ณผ reporting checklist (ยง5, ๋ถ€๋ก .1)

์ €์ž๋“ค์˜ ๊ฒฐ๋ก ์€ ์ด๋ ‡๋‹ค. teacher๊ฐ€ ์ ‘์ด‰ํ•˜๋Š” ๊ณณ์— ๋‘” ์†Œ์ˆ˜์˜ taxel์ด uniform์œผ๋กœ ๋‘” ํ›จ์”ฌ ํฐ ์ง‘ํ•ฉ์„ ๋งž์ท„๊ณ , ๋“œ๋ฌผ๊ฒŒ ์ ‘์ด‰ํ•˜๋Š” ๊ณณ์˜ taxel์€ ์“ธ ๋งŒํ•œ ์ •๋ณด๋ฅผ ์ฃผ์ง€ ์•Š์•˜์œผ๋ฉฐ, 2000๊ฐœ๋ฅผ ๋‹ค ์“ฐ๋ฉด layout์€ CNN์—์„œ๋„ ์„ฑ๋Šฅ์„ ๋ฐ”๊พธ์ง€ ์•Š์•˜๋‹ค. ์„ธ ๊ฒฐ๊ณผ์˜ ์„ค๋ช…์€ ํ•˜๋‚˜๋‹ค. ์ด‰๊ฐ ์‹ ํ˜ธ๋Š” ์ ‘์ด‰์ด ์ผ์–ด๋‚˜๋Š” ๊ณณ์—์„œ๋งŒ ์ •๋ณด๋ฅผ ์ค€๋‹ค. ๊ทธ๋Ÿฌ๋ฏ€๋กœ ์ด‰๊ฐ์— ๋Œ€ํ•œ ์งˆ๋ฌธ์€ โ€œ๋ช‡ ๊ฐœโ€๊ฐ€ ์•„๋‹ˆ๋ผ โ€œ์–ด๋А ๊ฒƒโ€์ด๊ณ , ๊ทธ ๋‹ต์€ ์ด‰๊ฐ policy๋ฅผ ํ•™์Šตํ•˜๊ธฐ ์ „์— teacher์˜ ์ ‘์ด‰ ํ†ต๊ณ„์—์„œ ์ฝ์„ ์ˆ˜ ์žˆ๋‹ค. uniform placement๋Š” ์ด๋ ‡๊ฒŒ โ€œ์ˆจ์€ ๊ธฐ๋ณธ๊ฐ’โ€์—์„œ โ€œ์ธก์ • ๊ฐ€๋Šฅํ•œ ์„ค๊ณ„ ์„ ํƒโ€์œผ๋กœ ๋ฐ”๋€๋‹ค. ์ €์ž๋“ค์€ teacher ๋Œ€์‹  ๋‹ค๋ฅธ policy์˜ rollout์ด๋‚˜ ์‚ฌ๋žŒ ์‹œ์—ฐ์—์„œ ๊ธฐ๋กํ•œ ์ ‘์ด‰ ์œ„์น˜๋กœ๋„ ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ taxel์„ ๊ณ ๋ฅผ ์ˆ˜ ์žˆ๋‹ค๊ณ  ๋ง๋ถ™์ด๊ณ , ๋ชจ๋“  ๊ฒฐ๊ณผ๊ฐ€ ์‹œ๋ฎฌ๋ ˆ์ด์…˜์ด๋ฉฐ ์„ผ์„œ noise์™€ drift์˜ ํšจ๊ณผ๋Š” ์‹œํ—˜ํ•˜์ง€ ์•Š์•˜๋‹ค๊ณ  ๋ฐํžŒ๋‹ค.

๊ทธ๋ฃน ์„ค๊ณ„ ์„ ํƒ ์ด ๋…ผ๋ฌธ์˜ ์„ค์ • (XHand1 In-Palm Rotation ๊ธฐ์ค€)
Sensing Candidate set N=2000, ์“ฐ๋Š” ๋งํฌ ์œ„ Poisson-disk sampling
Taxel signal ๋งํฌ ์ขŒํ‘œ๊ณ„์˜ 6-D ํž˜ยทํ† ํฌ
Selection (RQ1) Budgetโ€  B\in\{10,15,25,50,75,100,300,2000\}
Placement ruleโ€  Uniform, Top-contact, Contact-cover, Link-proportional, Least-contact
Contact statistics \pi_T rollout (์ด‰๊ฐ ์ž…๋ ฅ ์—†์Œ)
Representation (RQ2) Tensor size N taxel๋กœ ๊ณ ์ •, zero padding (Table 4)
Orderingโ€  Random, Band, Grid
Encoderโ€  LSTM, CNN, Link-wise MLP
Evaluation Protocol 1024 ์—ํ”ผ์†Œ๋“œ, ํ‰๊ท ๊ณผ ํ‘œ์ค€ํŽธ์ฐจ
Metric ์„ฑ๊ณต๋ฅ (In-Palm Rotation), ์™ธ๋ž€ ํ•˜ held fraction(Grasp-and-Lift)

Table 5. ์ด‰๊ฐ ๊ด€์ธก์˜ reporting checklist. โ€ ๋Š” ์ด ์—ฐ๊ตฌ์—์„œ ๋ฐ”๊พผ ํ•ญ๋ชฉ.

์ €์ž๋“ค์€ ์ด์ „ ์—ฐ๊ตฌ๋“ค์ด ์ด ์„ ํƒ์„ ์„œ๋กœ ๋‹ค๋ฅด๊ฒŒ ํ•˜๊ณ  ์ผ๋ถ€๋Š” ๋ณด๊ณ ์กฐ์ฐจ ํ•˜์ง€ ์•Š๋Š”๋‹ค๊ณ  ์ง€์ ํ•˜๋ฉฐ, ์ด ํ‘œ๋ฅผ ๋„ค ๊ทธ๋ฃน์œผ๋กœ ์ •๋ฆฌํ•œ๋‹ค. ์˜๋„๋Š” ์ข‹๋‹ค. ๋‹ค๋งŒ ์ด checklist์—๋Š” ์ด ๋…ผ๋ฌธ ์ž์‹ ์˜ ๊ฒฐ๊ณผ๋ฅผ ์ขŒ์šฐํ•œ ๋ช‡ ๊ฐ€์ง€๊ฐ€ ๋น ์ ธ ์žˆ๋‹ค(์•„๋ž˜ ์•ฝ์  ์ ˆ).

์„ค๊ณ„ ๋ ˆ์Šจ: ๋…ผ๋ฌธ์ด ๋งํ•œ ๊ฒƒ๊ณผ ํ‘œ๊ฐ€ ๋ฐ›์น˜๋Š” ๋ฒ”์œ„

์ €์ž์˜ ๋ ˆ์Šจ (ยง1) ํ‘œ๊ฐ€ ๋ฐ›์น˜๋Š” ๋ฒ”์œ„ ์‹ค๋ฌด๋กœ ์˜ฎ๊ธธ ๋•Œ
๋ช‡ ๊ฐœ๋ณด๋‹ค ์–ด๋А taxel์„ ๊ฐ์ง€ํ•˜๋А๋ƒ๊ฐ€ ์ค‘์š”ํ•˜๋‹ค XHand1 In-Palm Rotation์—์„œ ๊ฐ•ํ•˜๊ฒŒ(Table 1), Allegro Grasp-and-Lift์—์„œ Contact-cover์— ํ•œํ•ด ์•ฝํ•˜๊ฒŒ(Table 2) ํ•™์Šต ์ „์— teacher ์ ‘์ด‰ ํ†ต๊ณ„๋ถ€ํ„ฐ ๋ฝ‘์•„ ๋ณธ๋‹ค. ํ›„๋ณด์˜ ๋ช‡ %๊ฐ€ ํ•œ ๋ฒˆ๋„ ์•ˆ ๋‹ฟ๋Š”์ง€(XHand1์€ 44%)๊ฐ€ ์ฒซ ์ง„๋‹จ๊ฐ’์ด๋‹ค
๋“œ๋ฌผ๊ฒŒ ๋‹ฟ๋Š” taxel์€ ์ด‰๊ฐ์ด ์—†๋Š” ๊ฒƒ๊ณผ ๊ฐ™๋‹ค Least-contactยทUniform ์ €budget = 0.0%(Table 1), Contact Events 0.1~0.2%(Table 10) ์„ฑ๊ณต๋ฅ  ๋Œ€์‹  โ€œContact Eventsโ€ ๊ฐ™์€ coverage ์ง€ํ‘œ๋กœ placement๋ฅผ ๋จผ์ € ๊ฑธ๋Ÿฌ๋‚ผ ์ˆ˜ ์žˆ๋‹ค
encoder๋Š” tensor ์•ˆ ์œ„์น˜๊ฐ€ ์•„๋‹ˆ๋ผ ์–ด๋А taxel์ด ๋ณด๊ณ ํ–ˆ๋Š”์ง€๋ฅผ ๋ณด์กดํ•ด์•ผ ํ•œ๋‹ค 2000๊ฐœ ์ „๋ถ€, In-Palm Rotation์—์„œ๋งŒ(Table 3) ๊ฒฉ์ž ๋งคํ•‘์— ๊ณต์„ ๋“ค์ด๊ธฐ๋ณด๋‹ค taxel ์ •์ฒด์„ฑ์„ ์ง€์šฐ๋Š” pooling(๋งํฌ ํ‰๊ท  ๋“ฑ)์„ ํ”ผํ•œ๋‹ค
(Grasp-and-Lift์—์„œ ๋“œ๋Ÿฌ๋‚œ ๊ฒƒ) ์ž์ฃผ ๋‹ฟ๋Š” ๊ณณ์— ๋ชฐ์ง€ ๋ง๊ณ  ๋‹ฟ๋Š” ๋งํฌ๋ฅผ ๋ฎ์–ด๋ผ Top-contact ์‹คํŒจ, Contact-cover ์šฐ์œ„(Table 2, Table 12) ๊ทœ์น™์˜ ๊ธฐ๋ณธ๊ฐ’์€ Top-contact๋ณด๋‹ค Contact-cover๊ฐ€ ์•ˆ์ „ํ•˜๋‹ค

๋น„ํŒ์ ์œผ๋กœ ๋ณด๋ฉด

๊ฐ•์ 

  • ์งˆ๋ฌธ์„ ์ •ํ™•ํžˆ ์ชผ๊ฐฐ๋‹ค. ๊ฐ™์€ ํ›„๋ณด ์ง‘ํ•ฉ, ๊ฐ™์€ budget, nested ์„ ํƒ, ๊ฐ™์€ teacher ์ ‘์ด‰ ํ†ต๊ณ„๋ผ๋Š” ์žฅ์น˜๋กœ โ€œ๋ช‡ ๊ฐœโ€์™€ โ€œ์–ด๋А ๊ฒƒโ€์„ ์ฒ˜์Œ์œผ๋กœ ๋ถ„๋ฆฌํ–ˆ๋‹ค. Tactile Genesis์˜ fingertip vs ์† ์ „์ฒด ๋น„๊ต๊ฐ€ ์„ผ์„œ ์ˆ˜์™€ ์œ„์น˜๋ฅผ ํ•จ๊ป˜ ๋ฐ”๊ฟจ๋˜ ๊ฒƒ๊ณผ ๋Œ€๋น„๋œ๋‹ค. teacher๊ฐ€ ์ด‰๊ฐ์„ ๋ฐ›์ง€ ์•Š๊ฒŒ ํ•ด โ€œplacement๊ฐ€ teacher์˜ ์ ‘์ด‰์„ ๋ฐ”๊พธ๊ณ , ๊ทธ ์ ‘์ด‰์ด ๋‹ค์‹œ placement๋ฅผ ์ •ํ•˜๋Š”โ€ ์ˆœํ™˜์„ ๋Š์€ ๊ฒƒ๋„ ๊น”๋”ํ•˜๋‹ค.
  • ์Œ์„ฑ ๋Œ€์กฐ๊ตฐ์ด ์„ค๊ณ„์— ๋“ค์–ด ์žˆ๋‹ค. Least-contact์™€ Per-link force๊ฐ€ โ€œtaxel์ด ์žˆ์–ด๋„ ์ ‘์ด‰์„ ๋ชป ๋ณด๋ฉด ์—†๋Š” ๊ฒƒ๊ณผ ๊ฐ™๋‹คโ€์™€ โ€œํ•ด์ƒ๋„ ์—†๋Š” ์ง‘๊ณ„๋Š” ๋ถ€์กฑํ•˜๋‹คโ€๋ฅผ ๊ฐ๊ฐ ์ง์ ‘ ๋ณด์—ฌ ์ค€๋‹ค. ๋Œ€์กฐ๊ตฐ ์—†์ด โ€œ์„ ํƒ์ด ์ค‘์š”ํ•˜๋‹คโ€๋งŒ ์ฃผ์žฅํ–ˆ๋‹ค๋ฉด ํ›จ์”ฌ ์•ฝํ–ˆ์„ ๊ฒƒ์ด๋‹ค.
  • ๊ฒฐ๊ณผ๋ฅผ ๋ฉ”์ปค๋‹ˆ์ฆ˜ ์ง€ํ‘œ๋กœ ์„ค๋ช…ํ•œ๋‹ค. ์„ฑ๊ณต๋ฅ  ํ‘œ๋งŒ ๋‚ด์ง€ ์•Š๊ณ  Contact EventsยทTaxel ReadingsยทRedundancy(Table 10), ์—ํ”ผ์†Œ๋“œ ๊ฒฐ๊ณผ ๋ถ„ํ•ด(Table 11), ์ ‘์ด‰ ์ง‘์ค‘๋„(Table 12)๋ฅผ ๋ถ™์—ฌ โ€œ์™œโ€๋ฅผ ๋”ฐ๋ผ๊ฐˆ ์ˆ˜ ์žˆ๊ฒŒ ํ–ˆ๋‹ค. ํŠนํžˆ Contact Events๊ฐ€ ์„ฑ๊ณต๋ฅ ์„ ์ž˜ ๋”ฐ๋ผ๊ฐ„๋‹ค๋Š” ๊ด€์ฐฐ์€ ํ•™์Šต ์—†์ด placement๋ฅผ ๊ฑฐ๋ฅด๋Š” ๊ฐ’์‹ผ ์ง„๋‹จ์œผ๋กœ ์“ธ ์ˆ˜ ์žˆ๋‹ค.
  • ์ž๊ธฐ ๊ฒฐ๊ณผ์˜ ๋ฐ˜๋ก€๋ฅผ ์ˆจ๊ธฐ์ง€ ์•Š๋Š”๋‹ค. Grasp-and-Lift์—์„œ Top-contact๊ฐ€ ์‹คํŒจํ•˜๊ณ  ๊ฒฝํ–ฅ์ด ๋ฐ˜๋Œ€๋ผ๋Š” ์ , 2000๊ฐœ๋ฅผ ๋‹ค ์จ๋„ teacher์™€ ๊ฒฉ์ฐจ๊ฐ€ ํฌ๋‹ค๋Š” ์ , ๊ตํ›ˆ์ด ํ•œ ์†ยทํ•œ ๊ณผ์ œ์—์„œ ๋‚˜์™”๋‹ค๋Š” ์ ์„ ๋ณธ๋ฌธ๊ณผ ๊ฒฐ๋ก ์— ์ง์ ‘ ์ ์—ˆ๋‹ค. Link-wise MLP์˜ ์„ธ ์นธ ์ฐจ์ด๊ฐ€ seed ๋ถ„์‚ฐ์ด๋ผ๊ณ  ์Šค์Šค๋กœ ๋ฐํžŒ ๊ฒƒ๋„ ๋“œ๋ฌธ ์ •์งํ•จ์ด๋‹ค.
  • ์žฌํ˜„ ์ •๋ณด๊ฐ€ ๋น„๊ต์  ์ถฉ์‹คํ•˜๋‹ค. ๋ณด์ƒ ํ•ญ๋ชฉ๊ณผ ๊ฐ€์ค‘์น˜(Table 6ยท7), PPOยทDAgger ํ•˜์ดํผํŒŒ๋ผ๋ฏธํ„ฐ(Table 8), encoder ๊ตฌ์กฐ(Table 9), ์™ธ๋ž€ ์‚ฌ์–‘, tie-breaking ๊ทœ์น™๊นŒ์ง€ ์ ์—ˆ๋‹ค.

์•ฝ์ ยทํ•œ๊ณ„

  • ํ—ค๋“œ๋ผ์ธ ์ฃผ์žฅ์˜ ์ •๋Ÿ‰ ๊ทผ๊ฑฐ๊ฐ€ ํ•œ ์†ยทํ•œ ๊ณผ์ œ์— ์žˆ๋‹ค. ์œ„ ๊ทผ๊ฑฐ ์กฐ๊ฑด ํ‘œ์—์„œ ๋ณด๋“ฏ โ€œ์ž‘์€ ์„ ํƒ ์ง‘ํ•ฉ์ด ํ›จ์”ฌ ํฐ uniform ์ง‘ํ•ฉ์„ ๋งž์ถ˜๋‹คโ€๊ฐ€ ๊นจ๋—ํ•˜๊ฒŒ ์„ฑ๋ฆฝํ•˜๋Š” ๊ณณ์€ XHand1 In-Palm Rotation๋ฟ์ด๊ณ , Sharpa Grasp-and-Lift์—์„œ๋Š” ์„ฑ๋ฆฝํ•˜์ง€ ์•Š๋Š”๋‹ค. ์ €์ž๋„ ๊ฒฐ๋ก ์—์„œ ์ด๋ฅผ ์ธ์ •ํ•˜์ง€๋งŒ ์ดˆ๋ก์€ โ€œthree robot handsโ€๋กœ ์‹œ์ž‘ํ•ด โ€œUniform placement therefore overestimatesโ€ฆโ€๋กœ ์ผ๋ฐ˜ํ™”ํ•œ๋‹ค. ํ‘œํ˜„ ์‹คํ—˜(RQ2)๋„ XHand1 In-Palm Rotation ํ•œ ์กฐ๊ฑด์—์„œ๋งŒ ํ–ˆ๋‹ค.
  • Top-contact๋Š” ๋‘ ๊ณผ์ œ์—์„œ ๊ฐ™์€ ๊ทœ์น™์ด ์•„๋‹ˆ๋‹ค. ๋ถ€๋ก .5๋Š” Top-contactยทContact-coverยทLeast-contact๊ฐ€ โ€œFor In-Palm Rotationโ€ ๋ชจ๋“  ๋งํฌ์— ์ตœ์†Œ 1๊ฐœ์”ฉ taxel์„ ๋ฐฐ์ •ํ•œ๋‹ค๊ณ  ์ ๋Š”๋‹ค. Grasp-and-Lift์— ๋Œ€ํ•ด์„œ๋Š” ์ด ๊ทœ์น™์ด ์–ธ๊ธ‰๋˜์ง€ ์•Š๋Š”๋‹ค. ๊ทธ๋ ‡๋‹ค๋ฉด In-Palm Rotation์˜ Top-contact๋Š” ์‚ฌ์‹ค์ƒ โ€œ๋งํฌ coverage ๋ณด์žฅ + ์ƒ์œ„ ์ ‘์ด‰โ€์ด๊ณ , Grasp-and-Lift์˜ Top-contact๋Š” ์ˆœ์ˆ˜ํ•œ ์ƒ์œ„ ์ ‘์ด‰์ด๋‹ค. ์ €์ž๋“ค์ด Grasp-and-Lift์—์„œ Top-contact๊ฐ€ ์‹คํŒจํ•œ ์ด์œ ๋กœ โ€œ๋งํฌ๋ฅผ ๋ฎ๋Š” ๊ฒƒ์ด ์ค‘์š”ํ•˜๋‹คโ€๋ฅผ ๋“œ๋Š”๋ฐ, ๋ฐ”๋กœ ๊ทธ ์ฐจ์ด๊ฐ€ ๊ทœ์น™ ์ •์˜์—์„œ ์ƒ๊ฒผ์„ ์ˆ˜ ์žˆ๋‹ค. In-Palm Rotation์˜ Top-contact ์„ฑ๊ณต ์ค‘ ์–ผ๋งˆ๊ฐ€ coverage ๋ณด์žฅ ๋•์ธ์ง€๋Š” ๋ถ„๋ฆฌ๋˜์ง€ ์•Š์•˜๋‹ค. ๊ฐ™์€ ๋ฌธ๋‹จ์—์„œ zero-padding๋„ โ€œFor In-Palm Rotationโ€์œผ๋กœ ํ•œ์ •๋ผ ์žˆ์–ด, Grasp-and-Lift์˜ ์ž…๋ ฅ ์ฐจ์›์ด budget๋งˆ๋‹ค ๊ฐ™์€์ง€๋Š” ์ ํ˜€ ์žˆ์ง€ ์•Š๋‹ค.
  • 3 seeds์˜ ์žก์Œ ์ˆ˜์ค€์ด ์ผ๋ถ€ ๋น„๊ต๋ณด๋‹ค ํฌ๋‹ค. ์ €์ž๊ฐ€ ์Šค์Šค๋กœ ๋ฐํžŒ ๊ธฐ์ค€์ ์ด ์žˆ๋‹ค. Link-wise MLP์˜ ์„ธ layout์€ ๊ฐ™์€ ์ž…๋ ฅ์ธ๋ฐ ํ‰๊ท ์ด 18.6 ยฑ 1.1๊ณผ 33.2 ยฑ 8.6์œผ๋กœ ๊ฐˆ๋ฆฐ๋‹ค. ๊ฐ™์€ ๋ฐฉ์‹์œผ๋กœ Welch t๋ฅผ ๊ณ„์‚ฐํ•˜๋ฉด ์ด โ€œ๊ฐ€์งœ ์ฐจ์ดโ€๋Š” t โ‰ˆ 2.9์ด๊ณ , CNN Band(64.3) vs Grid(61.3)๋Š” t โ‰ˆ 3.1๋กœ ๊ฑฐ์˜ ๊ฐ™์€ ํฌ๊ธฐ๋‹ค(ํ•„์ž ๊ณ„์‚ฐ, ๋ณด๊ณ ๋œ std๋ฅผ ํ‘œ๋ณธ ํ‘œ์ค€ํŽธ์ฐจ๋กœ ๊ฐ€์ •). ๊ทธ๋Ÿฌ๋‹ˆ Table 3์˜ layout ๊ฐ„ ์ฐจ์ด๋Š” ์žก์Œ๊ณผ ๊ตฌ๋ถ„๋˜์ง€ ์•Š๊ณ , ์ด๋Š” โ€œlayout์ด ์ƒ๊ด€์—†๋‹คโ€๋Š” ์ €์ž ๊ฒฐ๋ก ์—๋Š” ์˜คํžˆ๋ ค ๋ถ€ํ•ฉํ•œ๋‹ค. ๋ฌธ์ œ๋Š” Table 2๋‹ค. Sharpa B=50์˜ Contact-cover 38.9 ยฑ 15.7 vs Uniform 24.0 ยฑ 0.8์€ t โ‰ˆ 1.6, Allegro B=300์˜ 72.2 ยฑ 2.1 vs 57.0 ยฑ 11.9๋Š” t โ‰ˆ 2.2๋กœ, ์œ„ ๊ธฐ์ค€์ ๋ณด๋‹ค๋„ ์ž‘๋‹ค. Grasp-and-Lift์—์„œ ๊ฐ•๊ฑดํ•ด ๋ณด์ด๋Š” ์ฐจ์ด๋Š” Sharpa B=300์˜ Contact-coverยทLink-prop. vs Uniform(t โ‰ˆ 4.9, 9.8)๊ณผ Top-contact์˜ ์—ด์„ธ(t โ‰ˆ โˆ’6.4) ์ •๋„๋‹ค. ๋ฐ˜๋ฉด In-Palm Rotation์˜ ์ €budget ๊ฒฉ์ฐจ(์˜ˆ: B=50 Uniform 27.5 vs Top-contact 53.3, t โ‰ˆ โˆ’11)๋Š” ์žก์Œ๋ณด๋‹ค ํ›จ์”ฌ ์ปค์„œ ํ•ต์‹ฌ ์ฃผ์žฅ์€ ์ด ๋น„ํŒ์„ ๊ฒฌ๋”˜๋‹ค. ํ‘œ์ค€ํŽธ์ฐจ๊ฐ€ ํ‘œ๋ณธ์ธ์ง€ ๋ชจ์ง‘๋‹จ์ธ์ง€, 1024 ์—ํ”ผ์†Œ๋“œ๊ฐ€ seed๋งˆ๋‹ค์ธ์ง€ ํ•ฉ๊ณ„์ธ์ง€๋„ ์ ํ˜€ ์žˆ์ง€ ์•Š๋‹ค.
  • teacher๊ฐ€ ํ•˜๋‚˜๋‹ค. ๋ชจ๋“  ํ•™์ƒ์€ ๊ฐ™์€ \pi_T ํ•˜๋‚˜์—์„œ distill๋˜๊ณ , ์ ‘์ด‰ ํ†ต๊ณ„๋„ ๊ทธ teacher ํ•˜๋‚˜์—์„œ ๋‚˜์˜จ๋‹ค. teacher ์„ฑ๋Šฅ(93.8ยท96.0ยท93.0)์—๋Š” ๋ถ„์‚ฐ์ด ์—†๋‹ค. ๋‹ค๋ฅธ seed์˜ teacher๊ฐ€ ๋‹ค๋ฅธ ์ ‘์ด‰ ์ „๋žต(๋‹ค๋ฅธ ์†๊ฐ€๋ฝ ์‚ฌ์šฉ)์„ ํ•™์Šตํ–ˆ๋‹ค๋ฉด placement๋Š” ๋‹ฌ๋ผ์ง€๊ณ , ํ•œ teacher์˜ ํ†ต๊ณ„๋กœ ๊ณ ๋ฅธ taxel์ด ๋‹ค๋ฅธ teacher์—์„œ distillํ•œ ํ•™์ƒ์—๊ฒŒ ์–ผ๋งˆ๋‚˜ ์˜ฎ๊ฒจ์ง€๋Š”์ง€๋Š” ์‹œํ—˜๋˜์ง€ ์•Š์•˜๋‹ค. ์ €์ž๊ฐ€ ๊ฒฐ๋ก ์—์„œ ์ œ์•ˆํ•˜๋Š” โ€œ๋‹ค๋ฅธ policy๋‚˜ ์‚ฌ๋žŒ ์‹œ์—ฐ์˜ ์ ‘์ด‰ ํ†ต๊ณ„๋กœ ๊ณ ๋ฅด๊ธฐโ€๊ฐ€ ๋ฐ”๋กœ ์ด ์ „์ด ๋ฌธ์ œ์ธ๋ฐ, ๊ทธ๊ฒƒ๋„ ์‹คํ—˜์ด ์—†๋‹ค.
  • task-informed๋Š” ํ•™์ƒ์ด ํ‰๋‚ด ๋‚ผ ๋Œ€์ƒ์˜ ์ ‘์ด‰์„ ๋ฏธ๋ฆฌ ๋ณด๋Š” ๊ฒƒ์ด๋‹ค. ํ•™์ƒ์€ teacher์˜ ํ–‰๋™์„ ๊ทธ๋Œ€๋กœ ๋”ฐ๋ผ๊ฐ€๋„๋ก ํ•™์Šต๋˜๋ฏ€๋กœ ํ•™์ƒ์˜ ์ ‘์ด‰ ๋ถ„ํฌ๋Š” teacher์™€ ๋น„์Šทํ•ด์ง€๋„๋ก ์••๋ ฅ์„ ๋ฐ›๋Š”๋‹ค. ๊ทธ๋ž˜์„œ teacher ํ†ต๊ณ„๋กœ ๊ณ ๋ฅธ taxel์ด ํ•™์ƒ์—๊ฒŒ ์œ ๋ฆฌํ•œ ๊ฒƒ์€ ์–ด๋А ์ •๋„ ์˜ˆ์ •๋œ ๊ฒฐ๊ณผ๋‹ค. ์ด๊ฒƒ ์ž์ฒด๊ฐ€ ์ž˜๋ชป์€ ์•„๋‹ˆ์ง€๋งŒ, RL๋กœ ์ด‰๊ฐ policy๋ฅผ ์ง์ ‘ ํ•™์Šตํ•˜๊ฑฐ๋‚˜ ์ด‰๊ฐ์— ๋ฐ˜์‘ํ•ด ๋‹ค๋ฅธ ์ ‘์ด‰ ์ „๋žต์ด ๋‚˜์˜ค๋Š” ๊ฒฝ์šฐ(์˜ˆ: ํ•™์ƒ์ด teacher์™€ ๋‹ค๋ฅธ ์†๊ฐ€๋ฝ์„ ์“ฐ๊ฒŒ ๋˜๋Š” ๊ฒฝ์šฐ)์—๋Š” ๊ฐ™์€ ์ด๋“์ด ๋ณด์žฅ๋˜์ง€ ์•Š๋Š”๋‹ค. โ€œplacement๋Š” ํ•™์Šต ์ „์— ์ •ํ•  ์ˆ˜ ์žˆ๋‹คโ€๋Š” ๊ฒฐ๋ก ์€ ์ •ํ™•ํžˆ๋Š” โ€œdistillation ์„ค์ •์—์„œ๋Š” teacher๋ฅผ ํ•™์Šตํ•œ ๋’ค, ํ•™์ƒ์„ ํ•™์Šตํ•˜๊ธฐ ์ „์— ์ •ํ•  ์ˆ˜ ์žˆ๋‹คโ€๋‹ค.
  • Table 10์˜ Contact Events๋Š” Contact-cover์˜ ๋ชฉ์ ํ•จ์ˆ˜์™€ ๊ฑฐ์˜ ๊ฐ™๋‹ค. Contact-cover๋Š” โ€œ๊ฐ™์€ ๋งํฌ์—์„œ ์•„์ง ์•ˆ ์žกํžŒ ์ ‘์ด‰ ์ด๋ฒคํŠธ๋ฅผ ๊ฐ€์žฅ ๋งŽ์ด ์žก๋Š” taxelโ€์„ greedy๋กœ ๊ณ ๋ฅด๊ณ , Contact Events๋Š” โ€œ์ ‘์ด‰ํ•œ ๋งํฌ ์œ„ ์„ ํƒ taxel์ด ํ•˜๋‚˜๋ผ๋„ ์žก์€ ์ด๋ฒคํŠธ์˜ ๋น„์œจโ€์ด๋‹ค. Contact-cover๊ฐ€ ์ด ์—ด์—์„œ 88.8~100%์ธ ๊ฒƒ์€ ๊ตฌ์กฐ์ƒ ์˜ˆ์ƒ๋˜๋Š” ๊ฐ’์ด๊ณ , ๋…๋ฆฝ์ ์ธ ์ฆ๊ฑฐ๋ผ๊ธฐ๋ณด๋‹ค ์ตœ์ ํ™”๊ฐ€ ์ž˜ ๋๋‹ค๋Š” ํ™•์ธ์ด๋‹ค. ๋งˆ์ฐฌ๊ฐ€์ง€๋กœ Taxel Readings๋Š” Top-contact์˜ ์ •์˜(\lvert\mathcal{E}_i\rvert ์ƒ์œ„)์™€ ๊ฑฐ์˜ ๊ฐ™๋‹ค. ๋‘ ์ง€ํ‘œ๊ฐ€ ์„ฑ๊ณต๋ฅ ๊ณผ ๋งž๋ฌผ๋ฆฐ๋‹ค๋Š” ๊ด€์ฐฐ์€ ์œ ํšจํ•˜์ง€๋งŒ, ์–ด๋А ๊ทœ์น™์ด โ€œ๋” ์ž˜ ๋ณธ๋‹คโ€๋Š” ๊ทผ๊ฑฐ๋กœ๋Š” ์ˆœํ™˜์ ์ด๋‹ค.
  • Per-link force๋งŒ noise๋ฅผ ๋ฐ›์•˜๋‹ค. ๋ถ€๋ก .5๋Š” Per-link force์— 0.08 N Gaussian noise์™€ ํฌ๊ธฐ clipping์„ ์ ์šฉํ•œ๋‹ค๊ณ  ์ ๋Š”๋ฐ, taxel ์ธก์ • x_{i,t}์— ๋Œ€ํ•ด์„œ๋Š” noiseยทclipping์— ๋Œ€ํ•œ ์„œ์ˆ ์ด ์›๋ฌธ ์–ด๋””์—๋„ ์—†๋‹ค(์ „๋ฌธ ๊ฒ€์ƒ‰). taxel์ด noise ์—†๋Š” ์‹ ํ˜ธ๋ผ๋ฉด โ€œ๊ณต๊ฐ„ ํ•ด์ƒ๋„ ์—†๋Š” ์ง‘๊ณ„๋Š” ๊ฑฐ์˜ ๋„์›€์ด ์•ˆ ๋œ๋‹คโ€(4.4%)๋Š” ๋น„๊ต์— noise ์œ ๋ฌด๊ฐ€ ์„ž์ธ๋‹ค. ๋˜ํ•œ Per-link force์™€ Link-wise MLP๋Š” ๋‘˜ ๋‹ค taxel ์ •์ฒด์„ฑ์ด ์—†์ง€๋งŒ ์„ฑ๋Šฅ์ด 4.4%์™€ 18.6~33.2%๋กœ ํฌ๊ฒŒ ๋‹ค๋ฅด๋‹ค. Link-wise MLP๋Š” ๋งํฌ ์‚ฌ์ด๊นŒ์ง€ ํ‰๊ท ์„ ๋‚ด๋ฏ€๋กœ ๋งํฌ ์ •์ฒด์„ฑ์กฐ์ฐจ Per-link force๋ณด๋‹ค ์ ๊ฒŒ ๋‚จ๋Š”๋‹ค. ๊ทธ๋Ÿฐ๋ฐ๋„ ๋” ๋†’๋‹ค๋Š” ๊ฒƒ์€ โ€œ๊ณต๊ฐ„ ํ•ด์ƒ๋„โ€ ๋ง๊ณ  ๋‹ค๋ฅธ ๊ฒƒ(์ ‘์ด‰ํ•œ taxel ์ˆ˜์˜ ๋ถ„ํฌ, ๊ฐœ๋ณ„ taxel์˜ ํ† ํฌ ๋“ฑ)์ด ์ •๋ณด๋ฅผ ์‹ค์–ด ๋‚˜๋ฅธ๋‹ค๋Š” ๋œป์ผ ์ˆ˜ ์žˆ๋Š”๋ฐ, ๋…ผ๋ฌธ์€ ์ด ์ฐจ์ด๋ฅผ ์„ค๋ช…ํ•˜์ง€ ์•Š๋Š”๋‹ค.
  • taxel ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ชจ๋ธ์ด ์„œ์ˆ ๋˜์ง€ ์•Š์•˜๋‹ค. โ€œ๊ฐ taxel์€ ๋งํฌ ์ขŒํ‘œ๊ณ„์˜ ํž˜๊ณผ ํ† ํฌ๋ฅผ ์žฐ๋‹คโ€๊ฐ€ ์ „๋ถ€๋‹ค. Genesis์—์„œ ๊ทธ ๊ฐ’์ด ์–ด๋–ป๊ฒŒ ๊ณ„์‚ฐ๋˜๋Š”์ง€(์ ‘์ด‰์ ์„ taxel์— ์–ด๋–ป๊ฒŒ ๋ฐฐ์ •ํ•˜๋Š”์ง€, ๋ฐ˜๊ฒฝ, ํ† ํฌ์˜ ๊ธฐ์ค€์ ), 4 timestep๋งˆ๋‹ค ํ‰๊ฐ€ํ•œ โ€œ์ ‘์ด‰โ€์˜ ๋ฌธํ„ฑ์ด ๋ฌด์—‡์ธ์ง€ ์ ํ˜€ ์žˆ์ง€ ์•Š๋‹ค. โ€œํ•œ ๋ฒˆ๋„ ๋‹ฟ์ง€ ์•Š๋Š” 44%โ€๋ผ๋Š” ํ•ต์‹ฌ ์ง„๋‹จ๊ฐ’์ด ์ด ๋ชจ๋ธ๊ณผ 4-step ์ƒ˜ํ”Œ๋ง์— ์ง์ ‘ ์˜์กดํ•œ๋‹ค.
  • checklist๊ฐ€ ์ž๊ธฐ ๊ฒฐ๊ณผ๋ฅผ ์ขŒ์šฐํ•œ ํ•ญ๋ชฉ์„ ๋นผ๋จน์—ˆ๋‹ค. Table 5์—๋Š” taxel noise ๋ชจ๋ธ, taxel ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๋ฐฉ์‹, ์ ‘์ด‰ ์ด๋ฒคํŠธ ์ƒ˜ํ”Œ๋ง ์ฃผ๊ธฐ, tie-breaking ๊ทœ์น™(B=50 Least-contact์˜ 23/50์„ ์ •ํ•จ), ๋งํฌ coverage ๋ณด์žฅ ๊ทœ์น™(๊ณผ์ œ๋ณ„๋กœ ๋‹ค๋ฆ„), ํ‰๊ฐ€ ์‹œ์ (๊ฐ™์€ iteration), seed ์ˆ˜์™€ teacher ์ˆ˜๊ฐ€ ์—†๋‹ค. ํŠนํžˆ ๋งํฌ coverage ๋ณด์žฅ์€ Top-contact์˜ ๊ณผ์ œ ๊ฐ„ ๋น„๊ต๋ฅผ ๋ฐ”๊ฟ€ ์ˆ˜ ์žˆ๋Š” ํ•ญ๋ชฉ์ด๋‹ค. โ€œminimal setโ€์ด๋ผ๋Š” ์˜๋„๋ฅผ ๊ฐ์•ˆํ•ด๋„, ์ด ๋…ผ๋ฌธ ์•ˆ์—์„œ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”๊พผ ์„ ํƒ์€ ๋“ค์–ด๊ฐ€์•ผ ํ•œ๋‹ค.
  • CNN์ด layout์— ๋‘”๊ฐํ•œ ์ด์œ ๊ฐ€ ๊ตฌ์กฐ์— ์žˆ์„ ์ˆ˜ ์žˆ๋‹ค. ์ด ๋…ผ๋ฌธ์˜ CNN์€ 3ร—3 conv ๋‘ ์ธต(receptive field 5ร—5) ๋’ค์— ๋ธ”๋ก๋ณ„ flatten๊ณผ ๋งํฌ๋ณ„ linear layer๊ฐ€ ๋ถ™๋Š”๋‹ค(Table 9). flatten ๋’ค์˜ linear layer๋Š” ์นธ๋งˆ๋‹ค ๋‹ค๋ฅธ ๊ฐ€์ค‘์น˜๋ฅผ ๊ฐ€์ง€๋ฏ€๋กœ, ๊ณ ์ •๋œ permutation์ด ์žˆ์–ด๋„ โ€œ์–ด๋А ์นธ์ด ์–ด๋А taxel์ธ๊ฐ€โ€๋ฅผ ํ•™์Šตํ•  ์ˆ˜ ์žˆ๋‹ค. ์ด๋ฏธ์ง€ ๋ถ„๋ฅ˜์—์„œ pixel shuffle์ด CNN์„ ๋ง๊ฐ€๋œจ๋ฆฌ๋Š” ๊ฒƒ์€ ๋Œ€๊ฐœ pooling์œผ๋กœ ์œ„์น˜ ์ •๋ณด๋ฅผ ๋ฒ„๋ฆฌ๊ณ  ๊ตญ์†Œ ํŒจํ„ด์— ์˜์กดํ•˜๊ธฐ ๋•Œ๋ฌธ์ด๋‹ค. ๊ทธ๋Ÿฌ๋‹ˆ โ€œCNN์—์„œ๋„ layout์ด ์ƒ๊ด€์—†๋‹คโ€๋Š” ๊ฒฐ๋ก ์€ โ€œflatten+linear๋กœ ์œ„์น˜๋ณ„ ๊ฐ€์ค‘์น˜๋ฅผ ๊ฐ€์ง„ ์–•์€ CNN์—์„œ๋Š”โ€์œผ๋กœ ์ฝ๋Š” ๊ฒƒ์ด ์ •ํ™•ํ•˜๋‹ค(ํ•„์ž ํ•ด์„). ์ €์ž์˜ ๋ ˆ์Šจ(โ€œ์ •์ฒด์„ฑ์„ ๋ณด์กดํ•˜๋ผโ€)๊ณผ๋Š” ์ผ๊ด€๋˜์ง€๋งŒ, global pooling์„ ์“ฐ๋Š” ๋” ๊นŠ์€ CNN์—์„œ๋„ ์„ฑ๋ฆฝํ•˜๋Š”์ง€๋Š” ์ด ์‹คํ—˜์ด ๋‹ตํ•˜์ง€ ์•Š๋Š”๋‹ค. ๋˜ ์ด ๋น„๊ต๋Š” 2000๊ฐœ ์ „๋ถ€, ์ฆ‰ ์ด˜์ด˜ํ•œ ํ‘œ๋ฉด์—์„œ๋งŒ ํ–ˆ๋‹ค. ํฌ์†Œํ•œ ์„ ํƒ ์ง‘ํ•ฉ์—์„œ Grid์˜ ์ด์›ƒ ๋ณด์กด์ด ์˜๋ฏธ๊ฐ€ ์žˆ๋Š”์ง€๋Š” ๋ฏธ์ง€์ˆ˜๋‹ค.
  • ์„ฑ๊ณต ๊ธฐ์ค€์ด ๋‚ฎ๊ณ  teacher์™€์˜ ๊ฒฉ์ฐจ๊ฐ€ ํฌ๋‹ค. In-Palm Rotation์˜ ์„ฑ๊ณต์€ โ€œ๋‚™ํ•˜ ์ „ 90๋„ ํšŒ์ „ ๋ชฉํ‘œ ํ•˜๋‚˜ ๋„๋‹ฌโ€์ด๋‹ค. 2000๊ฐœ๋ฅผ ๋‹ค ์จ๋„ 61.5%์ด๊ณ  teacher๋Š” 93.8%๋‹ค. ์ตœ๊ณ  ํ•™์ƒ(CNN Band 64.3%)๋„ teacher์™€ ์•ฝ 30%p ๋–จ์–ด์ ธ ์žˆ๋‹ค. ์ด‰๊ฐ ์„ค๊ณ„์˜ ์ฐจ์ด๊ฐ€ ์ด ํฐ distillation gap ์•„๋ž˜์˜ โ€œ๋‚จ์€ ๋ชซโ€์—์„œ๋งŒ ๋น„๊ต๋˜๋Š” ์…ˆ์ด๊ณ , ๋ชฉํ‘œ๋ฅผ ๋ช‡ ๊ฐœ ์—ฐ์† ๋„๋‹ฌํ–ˆ๋Š”์ง€ ๊ฐ™์€ ๋” ์ด˜์ด˜ํ•œ ์ง€ํ‘œ๋Š” ๋ณด๊ณ ๋˜์ง€ ์•Š๋Š”๋‹ค.
  • ๋น„์šฉ ํ‘œ์˜ ์ ˆ๊ฐ์€ ๋…ผ๋ฌธ์ด ์‹ค์ œ๋กœ ์–ป์€ ๊ฒƒ์ด ์•„๋‹ˆ๋‹ค. ๋ชจ๋“  policy๋ฅผ full observation์œผ๋กœ ํ•™์Šตํ–ˆ์œผ๋ฏ€๋กœ Table 4์˜ โ€œ์•ฝ 5๋ฐฐโ€๋Š” ๊ฐ€์ • ๋น„์šฉ์ด๋‹ค. ๊ทธ๋ฆฌ๊ณ  ๊ทธ์ค‘ ์ผ๋ถ€๋Š” padding ์ œ๊ฑฐ ํšจ๊ณผ๋‹ค(2000๊ฐœ ๋ฌดpadding ๋Œ€๋น„๋กœ๋Š” 3.5๋ฐฐ). ์‹ค์ œ ์†์—์„œ์˜ ๋ฐฐ์„ ยทreadout ๋Œ€์—ญํญ ์ ˆ๊ฐ์€ ๋™๊ธฐ๋กœ๋งŒ ์–ธ๊ธ‰๋˜๊ณ  ์ธก์ •๋˜์ง€ ์•Š์•˜๋‹ค.
  • ์ž‘์€ ์„œ์ˆ  ๋ถˆ์ผ์น˜. ๋ถ€๋ก .4๋Š” ์ž…๋ ฅ์„ โ€œSection III.E์˜ per-link blockโ€, ์ธก์ •์„ โ€œSection III.Dโ€๋กœ ๊ฐ€๋ฆฌํ‚ค์ง€๋งŒ ๋ณธ๋ฌธ ์ ˆ ๋ฒˆํ˜ธ๋Š” 3.5ยท3.4๋‹ค(LaTeX ์†Œ์Šค์— ํ•˜๋“œ์ฝ”๋”ฉ๋œ ์ฐธ์กฐ). Table 2 ์บก์…˜์—๋Š” seed ์ˆ˜๊ฐ€ ๋น ์ ธ ์žˆ๋‹ค. ๋ณธ๋ฌธ์˜ โ€œ๊ฐ€์žฅ ๋‚ฎ์€ budget์—์„œ ์ด‰๊ฐ ์—†์„ ๋•Œ๋ณด๋‹ค ๋” ์ž์ฃผ ๋–จ์–ด๋œจ๋ฆฐ๋‹คโ€๋Š” Table 11์ด B=25๋ถ€ํ„ฐ๋ผ B=10ยท15์— ๋Œ€ํ•ด์„œ๋Š” ํ™•์ธํ•  ์ˆ˜ ์—†๋‹ค.

๊ด€๋ จ ์—ฐ๊ตฌ์™€์˜ ์ž๋ฆฌ ๋งค๊น€

Tactile Genesis์™€์˜ ๊ด€๊ณ„. ์ด ๋…ผ๋ฌธ์€ Tactile Genesis์˜ ๊ณผ์ œ ์ •์˜(In-Palm Rotation, Grasp-and-Lift), ํ›„๋ณด ์ƒ์„ฑ(Poisson-disk sampling), ๊ด€์ธก ์‹ ํ˜ธ(per-taxel force/torque)๋ฅผ ๊ทธ๋Œ€๋กœ ๋ฌผ๋ ค๋ฐ›์€ ํ›„์† ํ†ต์ œ ์‹คํ—˜์ด๋‹ค. Tactile Genesis๋Š” โ€œ๋ฐฐ์น˜๊ฐ€ ์„ผ์„œ ์ข…๋ฅ˜๋ฅผ ์••๋„ํ•˜๊ณ , ํ•ด์ƒ๋„๋Š” coverage๋ณด๋‹ค ๋œ ์ค‘์š”ํ•˜๋‹คโ€๋Š” ๊ฒฐ๋ก ์„ ๋ƒˆ๋Š”๋ฐ, ๊ทธ ๋ฐฐ์น˜ ๋น„๊ต๋Š” uniform ์ƒ˜ํ”Œ๋ง ์œ„์—์„œ ์„ผ์„œ ์ˆ˜์™€ ์˜์—ญ์„ ํ•จ๊ป˜ ๋ฐ”๊พผ ๊ฒƒ์ด์—ˆ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ๊ทธ ๊ฒฐ๋ก ์„ ํ•œ ๋‹จ๊ณ„ ์ชผ๊ฐ ๋‹ค. ๊ฐ™์€ ์ˆ˜์—์„œ ์–ด๋А taxel์ด๋ƒ๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”๊พธ๊ณ , uniform์ด ๋งŽ์ด ํ•„์š”ํ•ด ๋ณด์ด๋Š” ๊ฒƒ์€ 0์˜ ์–‘ ๋•Œ๋ฌธ์ด๋ผ๋Š” ๊ฒƒ์ด๋‹ค. Tactile Genesis ๋ฆฌ๋ทฐ์—์„œ ์ •๋ฆฌํ•œ โ€œ์† ์ „์ฒด์— 200 taxel์ด๋ฉด ์ถฉ๋ถ„โ€์ด๋ผ๋Š” ์ˆซ์ž๋„ uniform placement ์œ„์˜ ๊ฐ’์ด๋ผ๋Š” ์ ์—์„œ, ์ด ๋…ผ๋ฌธ์€ ๊ทธ ์ˆซ์ž๋ฅผ โ€œ๊ณผ์ œ๊ฐ€ ์š”๊ตฌํ•˜๋Š” ์ˆ˜โ€๊ฐ€ ์•„๋‹ˆ๋ผ โ€œuniform์ด ์š”๊ตฌํ•˜๋Š” ์ˆ˜โ€๋กœ ๋‹ค์‹œ ์ฝ๊ฒŒ ๋งŒ๋“ ๋‹ค(ํ•„์ž ํ•ด์„, ๋‘ ๋…ผ๋ฌธ์˜ ์†ยท๊ณผ์ œยทbudget ๊ฒฉ์ž๊ฐ€ ๊ฐ™์ง€๋Š” ์•Š๋‹ค). ๋ฐ˜๋Œ€๋กœ Tactile Genesis๊ฐ€ ์‹ค๋ฌผ XHand1 ์ „์ด๋ฅผ ํ™•์ธํ•œ ๊ฒƒ๊ณผ ๋‹ฌ๋ฆฌ ์ด ๋…ผ๋ฌธ์€ ์‹ค๊ธฐ ๊ฒ€์ฆ์ด ์—†๋‹ค.

์˜์—ญ๋ณ„ ๊ธฐ์—ฌ ์—ฐ๊ตฌ. RoleOfTouch๋Š” Shadow Hand๋ฅผ 19๊ฐœ ํ•ด๋ถ€ํ•™ ์˜์—ญ์œผ๋กœ ๋‚˜๋ˆ„๊ณ  5๊ฐœ ์„ผ์„œ ์กฐํ•ฉ์„ ๋ฐ”๊ฟ” ์˜์—ญ ์ค‘์š”๋„ ์ง€๋„๋ฅผ ๋งŒ๋“ค์—ˆ๋‹ค. ์งˆ๋ฌธ์€ ๊ฐ™์ง€๋งŒ ๋‹จ์œ„๊ฐ€ ๋‹ค๋ฅด๋‹ค. ๊ทธ์ชฝ์€ ์‚ฌ๋žŒ์ด ์ •ํ•œ ์˜์—ญ๊ณผ ์กฐํ•ฉ ํƒ์ƒ‰, ์ด์ชฝ์€ 2000๊ฐœ ํ›„๋ณด์™€ teacher ์ ‘์ด‰ ํ†ต๊ณ„์— ๊ธฐ๋ฐ˜ํ•œ ๊ทœ์น™์ด๋‹ค. ๊ทธ์ชฝ ๋ฆฌ๋ทฐ์—์„œ ์ง€์ ํ•œ โ€œ๋ฐ˜๋ณต ์‹คํ–‰ยท๋ถ„์‚ฐ ์—†์Œโ€์„ ์ด ๋…ผ๋ฌธ์€ 3 seeds๋กœ ์ผ๋ถ€ ํ•ด์†Œํ–ˆ์ง€๋งŒ, ์œ„์—์„œ ๋ณธ ๊ฒƒ์ฒ˜๋Ÿผ 3 seeds๋กœ๋„ Grasp-and-Lift์˜ ์ผ๋ถ€ ๋น„๊ต๋Š” ์žก์Œ๊ณผ ๊ตฌ๋ถ„๋˜์ง€ ์•Š๋Š”๋‹ค. ๋‘ ๋…ผ๋ฌธ ๋ชจ๋‘ ์†๋ฐ”๋‹ฅ ์ชฝ ์ ‘์ด‰์˜ ์ค‘์š”์„ฑ์„ ๊ฐ€๋ฆฌํ‚จ๋‹ค๋Š” ์ ์€ ๊ฐ™๋‹ค(์ด ๋…ผ๋ฌธ Fig. 2์˜ XHand1 ์ ‘์ด‰ ๋นˆ๋„๋Š” ์†๋ฐ”๋‹ฅ์— ๋„“๊ฒŒ ํผ์ ธ ์žˆ๋‹ค). ์„ผ์„œ ์ˆ˜๋ฅผ ์ค„์ด๋ฉฐ ์„ฑ๋Šฅ์„ ์œ ์ง€ํ•œ Guo et al.(2024)๊ณผ ๊ฐ™์€ ์ˆ˜์—์„œ placement๊ฐ€ ๊ฒฐ๊ณผ๋ฅผ ๋ฐ”๊พผ๋‹ค๋Š” Birtalan & Koller(2026)๋„ ๊ฐ™์€ ๊ฐˆ๋ž˜๋‹ค.

์ด‰๊ฐ in-hand rotation๊ณผ distillation ๊ณ„๋ณด. privileged teacher๋ฅผ PPO๋กœ ํ•™์Šตํ•˜๊ณ  ๊ด€์ธก ๊ฐ€๋Šฅํ•œ ํ•™์ƒ์„ distillํ•˜๋Š” 2๋‹จ๊ณ„ ๊ตฌ์กฐ๋Š” Qi et al.์˜ RMA ๊ณ„์—ด in-hand rotation(2022)๊ณผ RotateIt(2023)์—์„œ ์™”๋‹ค. RotateIt์€ in-hand rotation์—์„œ ์ ‘์ด‰ ์œ„์น˜๊ฐ€ ์„ฑ๋Šฅ์„ ํฌ๊ฒŒ ์˜ฌ๋ฆฐ๋‹ค๋Š” ๊ทผ๊ฑฐ๋กœ ์ธ์šฉ๋œ๋‹ค. ์† ์ „์ฒด์— ๊น binary ์ ‘์ด‰ ์„ผ์„œ๋งŒ์œผ๋กœ ํšŒ์ „์„ ๋ฐฐ์šด Rotating without Seeing๋„ โ€œ๊ฐ€๋ ค์ง„ ์ƒํƒœ์˜ ์กฐ์ž‘โ€์„ ์ด‰๊ฐ์˜ ๋™๊ธฐ๋กœ ๋“œ๋Š” ์„ ํ–‰ ์—ฐ๊ตฌ๋กœ ์ธ์šฉ๋œ๋‹ค. ์ด ๋…ผ๋ฌธ์€ ๊ทธ ๊ณ„๋ณด์˜ ๊ด€์ธก ์„ค๊ณ„(์–ด๋””์— ๋ช‡ ๊ฐœ, ์–ด๋–ค tensor๋กœ)๋ฅผ ์ฒ˜์Œ์œผ๋กœ ํ†ต์ œ ๋ณ€์ˆ˜๋กœ ๋‹ค๋ฃฌ ์—ฐ๊ตฌ๋กœ ์ž๋ฆฌ๋งค๊น€ํ•  ์ˆ˜ ์žˆ๋‹ค. distillation ๊ธฐ๋ฐ˜ ์ด‰๊ฐ policy๋ฅผ ๋น„ํŒ์ ์œผ๋กœ ์ฝ์„ ๋•Œ์˜ ๊ณตํ†ต ์งˆ๋ฌธ(ํ•™์ƒ๊ณผ teacher์˜ ๊ฒฉ์ฐจ, teacher๊ฐ€ ๋ฌด์—‡์„ ๋ณด์•˜๋Š”๊ฐ€)์€ PTLD ๋ฆฌ๋ทฐ์—์„œ๋„ ๋‹ค๋ค˜๋Š”๋ฐ, ์ด ๋…ผ๋ฌธ์—์„œ๋„ teacherโ€“ํ•™์ƒ ๊ฒฉ์ฐจ(์•ฝ 30%p)๊ฐ€ ์ด‰๊ฐ ์„ค๊ณ„ ๋น„๊ต์˜ ์ƒํ•œ์œผ๋กœ ์ž‘๋™ํ•œ๋‹ค.

์š”์•ฝ

์ด ๋…ผ๋ฌธ์€ ์ด‰๊ฐ policy ์—ฐ๊ตฌ์—์„œ ๊ฒ€์ฆ ์—†์ด ์“ฐ์—ฌ ์˜จ uniform taxel placement๋ฅผ, teacherยทํ•™์Šต ์ ˆ์ฐจยทํ›„๋ณด 2000๊ฐœ๋ฅผ ๊ณ ์ •ํ•˜๊ณ  ์„ ํƒ ๊ทœ์น™๊ณผ ํ‘œํ˜„๋งŒ ๋ฐ”๊พธ๋Š” ํ†ต์ œ ํ”„๋กœํ† ์ฝœ๋กœ ์‹œํ—˜ํ–ˆ๋‹ค. XHand1 In-Palm Rotation(LSTM, 3 seeds)์—์„œ teacher ์ ‘์ด‰ ํ†ต๊ณ„๋กœ ๊ณ ๋ฅธ 25๊ฐœ๊ฐ€ uniform 100๊ฐœ์™€ ๊ฐ™์€ ์„ฑ๊ณต๋ฅ ์„ ๋ƒˆ๊ณ , uniform 10ยท15๊ฐœ์™€ ๋“œ๋ฌผ๊ฒŒ ๋‹ฟ๋Š” taxel์€ ์ด‰๊ฐ์ด ์—†๋Š” ๊ฒƒ๊ณผ ๊ฐ™์•˜๋‹ค. 2000๊ฐœ๋ฅผ ๋‹ค ์“ฐ๋ฉด tensor layout์€ CNN์—์„œ๋„ ์„ฑ๋Šฅ์„ ๋ฐ”๊พธ์ง€ ์•Š์•˜๊ณ , taxel ์ •์ฒด์„ฑ์„ ์ง€์šฐ๋Š” encoder๋งŒ ํฌ๊ฒŒ ๋–จ์–ด์กŒ๋‹ค. โ€œ์ด‰๊ฐ์€ ์ ‘์ด‰์ด ์žˆ๋Š” ๊ณณ์—์„œ๋งŒ ์ •๋ณด๋ฅผ ์ค€๋‹คโ€๋Š” ๋‹จ์ผ ์„ค๋ช…๊ณผ Contact Events ๊ฐ™์€ ์ง„๋‹จ ์ง€ํ‘œ, ์Œ์„ฑ ๋Œ€์กฐ๊ตฐ, ๊ทธ๋ฆฌ๊ณ  reporting checklist๋Š” ์ด ๋ถ„์•ผ์˜ ์‹คํ—˜ ์„ค๊ณ„๋ฅผ ํ•œ ๋‹จ๊ณ„ ์—„๋ฐ€ํ•˜๊ฒŒ ๋งŒ๋“ ๋‹ค.

ํ•œ๊ณ„๋Š” ๋ฒ”์œ„์™€ ๋น„๊ต์˜ ๋Œ€์นญ์„ฑ์— ์žˆ๋‹ค. ํ—ค๋“œ๋ผ์ธ ์ฃผ์žฅ์˜ ์ •๋Ÿ‰ ๊ทผ๊ฑฐ๋Š” ํ•œ ์†ยทํ•œ ๊ณผ์ œ์— ์žˆ๊ณ , Grasp-and-Lift์—์„œ๋Š” Top-contact๊ฐ€ ์‹คํŒจํ•˜๋ฉฐ Sharpa์—์„œ๋Š” โ€œ์ ์€ ์ˆ˜๋กœ ํฐ uniform์„ ๋งž์ถ˜๋‹คโ€๊ฐ€ ์„ฑ๋ฆฝํ•˜์ง€ ์•Š๋Š”๋‹ค. ๋งํฌ coverage ๋ณด์žฅ์ด In-Palm Rotation์—๋งŒ ์ ์šฉ๋ผ Top-contact์˜ ๊ณผ์ œ ๊ฐ„ ๋น„๊ต๊ฐ€ ๊ฐ™์€ ๊ทœ์น™์ด ์•„๋‹ˆ๊ณ , teacher๋Š” ํ•˜๋‚˜์ด๋ฉฐ, taxel์—๋Š” noise๊ฐ€ ๋ช…์‹œ๋˜์ง€ ์•Š์€ ๋ฐ˜๋ฉด Per-link force์—๋Š” noise๊ฐ€ ๋ถ™์—ˆ๋‹ค. 3 seeds์˜ ์žก์Œ์€ ์ €์ž ์Šค์Šค๋กœ ๋ณด์ธ Link-wise MLP์˜ 14.6%p ์ฐจ์ด๋งŒํผ ํฌ๋‹ค. ์—ฌ๋Ÿฌ teacher seed, Grasp-and-Lift์—์„œ์˜ coverage ๋ณด์žฅ ํ†ต์ผ, taxel noise ๋ชจ๋ธ, ์‹ค๊ธฐ ๊ฒ€์ฆ์ด ๋”ํ•ด์ง„๋‹ค๋ฉด โ€œwhich taxels, not how manyโ€๋Š” ์‹œ๋ฎฌ๋ ˆ์ด์…˜ ๊ด€์ฐฐ์„ ๋„˜์–ด ์ด‰๊ฐ ์† ์„ค๊ณ„์˜ ์ง€์นจ์ด ๋  ๊ฒƒ์ด๋‹ค.

์ด ๊ธ€์— ๋Œ€ํ•ด

  • ์ฝ์€ ๋ฒ”์œ„: arXiv v1(2611.10647v1)์˜ HTML๊ณผ LaTeX ์†Œ์Šค. ๋ณธ๋ฌธ๊ณผ ๋ถ€๋ก .1~.7์˜ ๋ชจ๋“  ํ‘œ(Table 1~12)๋ฅผ ํ™•์ธํ–ˆ๊ณ , ๊ฒฐ๊ณผ ํ‘œ๋Š” ์› ํ‘œ์˜ ์ˆซ์ž๋ฅผ ๊ทธ๋Œ€๋กœ ๋งˆํฌ๋‹ค์šด์œผ๋กœ ์˜ฎ๊ฒผ๋‹ค. ๊ทธ๋ฆผ 5์žฅ(Fig. 1~5)์€ ๋…ผ๋ฌธ์˜ ์ „์ฒด ๊ทธ๋ฆผ์ด๋‹ค.
  • ๊ณต๊ฐœ ๋ฒ”์œ„: ๋…ผ๋ฌธ๊ณผ LaTeX ์†Œ์Šค์— ์ฝ”๋“œยทํ”„๋กœ์ ํŠธ ํŽ˜์ด์ง€ยท์˜์ƒ ๋งํฌ๊ฐ€ ์—†๊ณ (์‹œ๋ฎฌ๋ ˆ์ดํ„ฐ Genesis์˜ ์ €์žฅ์†Œ ๋งํฌ๋งŒ ์žˆ๋‹ค), ๋ณ„๋„๋กœ ๊ณต๊ฐœ๋œ ์ฝ”๋“œ๋‚˜ ํ”„๋กœ์ ํŠธ ํŽ˜์ด์ง€๋„ ์ฐพ์ง€ ๋ชปํ–ˆ๋‹ค. ์ด ๊ธ€์˜ ๋ชจ๋“  ๊ฒฐ๊ณผ ์ˆ˜์น˜๋Š” ๋…ผ๋ฌธ์ด ๋ณด๊ณ ํ•œ ๊ฐ’์ด๋ฉฐ ์žฌํ˜„ ํ™•์ธ์€ ํ•˜์ง€ ์•Š์•˜๋‹ค.
  • ํ‘œ๊ธฐ: โ€œํ•„์ž ๊ณ„์‚ฐโ€์€ ๋…ผ๋ฌธ ์ˆ˜์น˜๋กœ ์ง์ ‘ ๊ณ„์‚ฐํ•œ ๊ฐ’(๋น„์œจ, ๋ฐฐ์ˆ˜, Welch t ํ†ต๊ณ„๋Ÿ‰)์ด๊ณ , Welch t๋Š” ๋ณด๊ณ ๋œ ํ‘œ์ค€ํŽธ์ฐจ๋ฅผ 3 seeds์˜ ํ‘œ๋ณธ ํ‘œ์ค€ํŽธ์ฐจ๋กœ ๊ฐ€์ •ํ•œ ๊ทผ์‚ฌ๋‹ค. โ€œํ•„์ž ํ•ด์„ยท์ถ”๋ก โ€์€ ๋…ผ๋ฌธ์ด ๋ช…์‹œํ•˜์ง€ ์•Š์€ ํ•ด์„์ด๋‹ค(Table 1์˜ layout์ด Grid๋ผ๋Š” ์ถ”๋ก , CNN ๊ตฌ์กฐ์— ๋Œ€ํ•œ ํ•ด์„, Tactile Genesis์˜ 200 taxel ์ˆ˜์น˜๋ฅผ ๋‹ค์‹œ ์ฝ๋Š” ํ•ด์„). ํ‘œ์˜ ๊ตต๊ฒŒยท๊ธฐ์šธ์ž„์€ ์› ํ‘œ์˜ ์ตœ๊ณ ยท2์œ„ ์ƒ‰ ๊ฐ•์กฐ๋ฅผ ์˜ฎ๊ธด ๊ฒƒ์ด๋‹ค.
  • ์ œ๋ชฉ: ๋…ผ๋ฌธ์— ์•ฝ์นญ์ด ์—†์–ด ์ œ๋ชฉ ์•ž๋ถ€๋ถ„์„ ๋”ฐ โ€œWhich Taxelsโ€๋กœ ๋ถˆ๋ €๋‹ค.

Copyright 2026, JungYeon Lee