Index
2026-08-21 — Experiment (중간 결과 — G3 진행 중)

E2E 실행 SR 결과 — VLM eval 답 재생: Stage B(접지·도달) → Stage C(VLA 실조작)

scene_bench | 353질의 × 5전략 × 2모델 + oracle-loc · 고유 플랜 1,216 · Stage C 81 플랜 실행 · VLM 재호출 0

TL;DR

0.83
oracle SR_B
0.24
oracle SR_C
0.17
P(lift|reached)
81
Stage C 플랜

1 무엇을 어떻게 쟀나

VLM eval(S3) 답 {category, room, surface}재생: 벤치마크 종료 상태(final_state.npz 25 pose + 로봇 base) 복원 → 답의 (room, surface)를 접지 v3(답 granularity까지 메모리, 그 가구 위 인스턴스 찾기는 하네스, 후보 ≤2) → 텔레포트 → Stage B reached(GT ≤0.85 m) → Stage C VLA(mixv2-20k, zed2 입력) "Pick up the {category}" 600스텝 × 조건부 재시도 2, 판정 GT body any-of lift(프로브 실행 전 생성). 집계는 질의 단위(기권·미접지·미도달 = 0), reached_only 질의는 Stage B가 최종.

접지 규칙 v1→v3 진화(oracle 0.63→0.82)와 Stage C 재시도·프로브 버그 수정(스모크에서 발견)은 claude/260821/exp-e2e_exec_stage_b.md, exp-e2e_exec_stage_c.md 참조.

2 셀별 SR_B / SR_C (질의 단위, n=gradable)

cellnabstainSR_BSR_CP(lift|reached, lifted-only)n_lift_eval
oracle-loc3480.000.8300.2390.166247
frames-uniform@gemini-3.6-flash3480.030.5890.1780.192177
rollforward-vis-pure@gemini-3.6-flash3480.020.6120.1670.171187
textlog@gemini-3.6-flash3480.380.4050.1090.149121
textlog@claude-sonnet-53480.410.3740.1010.144111
frames-uniform@claude-sonnet-53480.050.3620.0950.184114
rollforward-vis-pure@claude-sonnet-53480.050.3790.0950.161118
rollforward-v2@claude-sonnet-53480.320.2960.0780.14689
blind@gemini-3.6-flash3480.410.2160.0660.17563
rollforward-v2@gemini-3.6-flash3480.370.2440.0660.16274
blind@claude-sonnet-53480.530.1520.0320.12548

플랜 단위 Stage C 성공(소스별): blind@claude-sonnet-5 6/12, blind@gemini-3.6-flash 11/16, frames-uniform@claude-sonnet-5 21/28, frames-uniform@gemini-3.6-flash 34/52, oracle-loc 41/69, rollforward-v2@claude-sonnet-5 13/19, rollforward-v2@gemini-3.6-flash 12/22, rollforward-vis-pure@claude-sonnet-5 19/29, rollforward-vis-pure@gemini-3.6-flash 32/48, textlog@claude-sonnet-5 16/32, textlog@gemini-3.6-flash 18/34

3 층별 분해 (SR_B / SR_C)

질의 종류

cellfind_distractorfind_target
oracle-loc0.91 / 0.21 (n=173)0.75 / 0.27 (n=175)
frames-uniform@gemini-3.6-flash0.65 / 0.18 (n=165)0.57 / 0.19 (n=172)
rollforward-vis-pure@gemini-3.6-flash0.71 / 0.15 (n=165)0.55 / 0.19 (n=175)
textlog@gemini-3.6-flash0.62 / 0.14 (n=42)0.66 / 0.18 (n=174)
textlog@claude-sonnet-50.50 / 0.12 (n=32)0.66 / 0.18 (n=174)
frames-uniform@claude-sonnet-50.50 / 0.13 (n=160)0.27 / 0.07 (n=170)
rollforward-vis-pure@claude-sonnet-50.58 / 0.13 (n=157)0.24 / 0.07 (n=172)
rollforward-v2@claude-sonnet-50.54 / 0.10 (n=98)0.36 / 0.12 (n=139)
blind@gemini-3.6-flash0.42 / 0.11 (n=97)0.31 / 0.11 (n=110)
rollforward-v2@gemini-3.6-flash0.37 / 0.09 (n=67)0.39 / 0.11 (n=153)
blind@claude-sonnet-50.41 / 0.07 (n=76)0.25 / 0.07 (n=88)

채점 방식

cellliftedreached_only
oracle-loc0.84 / 0.14 (n=293)0.76 / 0.76 (n=55)
frames-uniform@gemini-3.6-flash0.63 / 0.12 (n=283)0.52 / 0.52 (n=54)
rollforward-vis-pure@gemini-3.6-flash0.66 / 0.11 (n=285)0.47 / 0.47 (n=55)
textlog@gemini-3.6-flash0.72 / 0.11 (n=168)0.42 / 0.42 (n=48)
textlog@claude-sonnet-50.70 / 0.10 (n=159)0.40 / 0.40 (n=47)
frames-uniform@claude-sonnet-50.41 / 0.08 (n=278)0.23 / 0.23 (n=52)
rollforward-vis-pure@claude-sonnet-50.43 / 0.07 (n=277)0.27 / 0.27 (n=52)
rollforward-v2@claude-sonnet-50.46 / 0.07 (n=194)0.33 / 0.33 (n=43)
blind@gemini-3.6-flash0.38 / 0.07 (n=166)0.29 / 0.29 (n=41)
rollforward-v2@gemini-3.6-flash0.43 / 0.07 (n=173)0.23 / 0.23 (n=47)
blind@claude-sonnet-50.35 / 0.04 (n=136)0.18 / 0.18 (n=28)

단일/다중

cellmultisingle
oracle-loc0.77 / 0.26 (n=35)0.84 / 0.24 (n=313)
frames-uniform@gemini-3.6-flash0.24 / 0.09 (n=33)0.65 / 0.19 (n=304)
rollforward-vis-pure@gemini-3.6-flash0.30 / 0.06 (n=33)0.66 / 0.18 (n=307)
textlog@gemini-3.6-flash0.63 / 0.22 (n=27)0.66 / 0.17 (n=189)
textlog@claude-sonnet-50.65 / 0.23 (n=26)0.63 / 0.16 (n=180)
frames-uniform@claude-sonnet-50.06 / 0.03 (n=34)0.42 / 0.11 (n=296)
rollforward-vis-pure@claude-sonnet-50.10 / 0.06 (n=31)0.43 / 0.10 (n=298)
rollforward-v2@claude-sonnet-50.19 / 0.12 (n=26)0.46 / 0.11 (n=211)
blind@gemini-3.6-flash0.14 / 0.05 (n=22)0.39 / 0.12 (n=185)
rollforward-v2@gemini-3.6-flash0.25 / 0.11 (n=28)0.41 / 0.10 (n=192)
blind@claude-sonnet-50.14 / 0.07 (n=14)0.34 / 0.07 (n=150)

바닥 여부

cellfloornot_floor
oracle-loc0.93 / 0.80 (n=30)0.82 / 0.19 (n=318)
frames-uniform@gemini-3.6-flash0.43 / 0.33 (n=30)0.63 / 0.17 (n=307)
rollforward-vis-pure@gemini-3.6-flash0.30 / 0.30 (n=30)0.66 / 0.16 (n=310)
textlog@gemini-3.6-flash0.40 / 0.33 (n=30)0.69 / 0.15 (n=186)
textlog@claude-sonnet-50.40 / 0.33 (n=30)0.67 / 0.14 (n=176)
frames-uniform@claude-sonnet-50.03 / 0.03 (n=30)0.42 / 0.11 (n=300)
rollforward-vis-pure@claude-sonnet-50.07 / 0.03 (n=29)0.43 / 0.11 (n=300)
rollforward-v2@claude-sonnet-50.24 / 0.16 (n=25)0.46 / 0.11 (n=212)
blind@gemini-3.6-flash0.10 / 0.10 (n=20)0.39 / 0.11 (n=187)
rollforward-v2@gemini-3.6-flash0.30 / 0.20 (n=30)0.40 / 0.09 (n=190)
blind@claude-sonnet-50.00 / 0.00 (n=12)0.35 / 0.07 (n=152)

쌍둥이 인접

cellFalseTrue
oracle-loc0.81 / 0.24 (n=304)0.98 / 0.25 (n=44)
frames-uniform@gemini-3.6-flash0.59 / 0.18 (n=295)0.76 / 0.19 (n=42)
rollforward-vis-pure@gemini-3.6-flash0.62 / 0.17 (n=298)0.67 / 0.19 (n=42)
textlog@gemini-3.6-flash0.62 / 0.18 (n=177)0.82 / 0.15 (n=39)
textlog@claude-sonnet-50.58 / 0.17 (n=169)0.86 / 0.16 (n=37)
frames-uniform@claude-sonnet-50.35 / 0.10 (n=287)0.58 / 0.12 (n=43)
rollforward-vis-pure@claude-sonnet-50.38 / 0.09 (n=287)0.52 / 0.14 (n=42)
rollforward-v2@claude-sonnet-50.41 / 0.10 (n=203)0.59 / 0.18 (n=34)
blind@gemini-3.6-flash0.34 / 0.11 (n=175)0.47 / 0.09 (n=32)
rollforward-v2@gemini-3.6-flash0.36 / 0.11 (n=180)0.53 / 0.10 (n=40)
blind@claude-sonnet-50.30 / 0.06 (n=132)0.41 / 0.09 (n=32)

4 실패 사례 (oracle-loc 기준 — 메모리가 아니라 하네스/정책 한계)

val_11__2 · Pick up the remote control I moved earlier · sources oracle-loc, textlog@gemini-3.6-flash, frames-uniform@gemini-3.6-flash…
FAIL · 476s · sel0_a0(remotecontrol_dbbd): t0:✗600, t1:✗600

첫 pick 키프레임 4장 (bench_exo_L)
val_144__0 · Pick up the apple I moved earlier · sources oracle-loc, blind@gemini-3.6-flash, textlog@gemini-3.6-flash…
FAIL · 494s · sel0_a0(apple_ff93afa9f54b): t0:✗600, t1:✗600

첫 pick 키프레임 4장 (bench_exo_L)
val_179__0 · Pick up all the ladles · sources oracle-loc
FAIL · 662s · sel0_a0(ladle_efa098c0c2ee): t0:✓300 | sel1_a0(ladle_efa098c0c2ee): t0:✗600 | sel1_a1(ladle_efa098c0c2ee): t0:✗600

첫 pick 키프레임 4장 (bench_exo_L)
val_101__1 · Pick up all the bowls I moved earlier · sources oracle-loc
FAIL · 626s · sel0_a0(bowl_46a21212675e4): t0:✓189 | sel1_a0(painting_2fadc8a08): t0:✗600 | sel1_a1(bowl_6befd62f08fd3): t0:✗600

첫 pick 키프레임 4장 (bench_exo_L)

성공 사례

val_101__1 · Pick up the remote control I moved earlier · sources oracle-loc, textlog@gemini-3.6-flash, rollforward-v2@gemini-3.6-flash…
SUCCESS · 318s · sel0_a0(remotecontrol_dbbd): t0:✗600 | sel0_a1(remotecontrol_dbbd): t0:✓118

마지막 성공 pick 키프레임 4장 (bench_exo_L)
val_125__0 · Pick up the fork · sources oracle-loc, frames-uniform@gemini-3.6-flash, rollforward-vis-pure@gemini-3.6-flash…
SUCCESS · 25s · sel0_a0(fork_a17d780ca6a79): t0:✓55

마지막 성공 pick 키프레임 4장 (bench_exo_L)
val_179__0 · Pick up the tissue box · sources oracle-loc, frames-uniform@gemini-3.6-flash, rollforward-vis-pure@gemini-3.6-flash…
SUCCESS · 136s · sel0_a0(tissuepaper_1f2829): t0:✓327

마지막 성공 pick 키프레임 4장 (bench_exo_L)

5 Takeaway / Next