Daily Progress — 2026-10-09

try-and-sucHigh

어제 한 일

시뮬 게이트 실험 통과: MuJoCo 퍽/공, 3차 누적 성공률 히스토리(VA) vs 기억 없음 — 학습 범위 87.1/87.5 vs 16.8/15.3 (+71.2%p), 보간 87.5/89.9 vs 0.0/0.0 (+88.7%p), 외삽 45.8/60.0 vs 2.9/1.7 (+50.6%p).
셔플 시 차이 +88.7 → +8.2%p, 히스토리 삭제 시 0.0%p → 격차는 히스토리 내용에서 발생. 영상만 쓰면 보간 59.0%, 영상+행동 88.7%.
기준선: PF 교사 85.1/86.5, IRP식 83.5/84.1, LLM 82.0/88.0(보간 @3). 외삽에서는 PF 80.8/78.2, IRP식 77.9/81.7로 우리(45.8/60.0)보다 높음 → 차별점 미입증.
code-reviewer 검토에서 blind sweep 인덱스 버그 발견·수정(결론 불변). 보간 띠 위치 재설계 후 데이터 재생성.
SLURM 약 20개 잡(gen 2, train 7, irp 2, eval 4 등), 대부분 COMPLETED. 의존성 걸린 첫 train 6개는 CANCELLED 후 재제출.

오늘 할 일

게이트 2단계: 연속 EE 속도 시퀀스 행동으로 확장, LLM/IRP 기준선과 재비교 (차별점 검증의 핵심).
외삽 개선(물성 범위 확대, DAgger식 재학습), 과제 5개 이상으로 task-holdout 재검증, 주 모델 3시드 반복.
Long-WAM 코드에서 과거 액션 컨텍스트 지원 여부 확인.

Long-WAM (LongLive)High

어제 한 일

RoboTwin 2.0 재현 스모크: 평가 job 181500 — lift_pot/click_bell/adjust_bottle 각 5/5 (15/15, 20분31초). 선행 181487(draccus 누락), 181491(train_transforms TypeError) 실패 후 수정.
학습 스모크 job 181499 (40 step, batch 16): loss 3.05 → 0.93, 5.7 s/step, 11분06초. 본 학습 6,119,051 샘플×5 epoch ≈ 29.9K update, 8 GPU 가정 시 6.3~15.8일.
text cache 누락 902,384개 중 job 181534(2 GPU)가 1시간25분 만에 COMPLETED. 번들 config 릴리스 버그 확인(우회 사본 사용).
git: 외부 커밋 1건(4ce82ab, arXiv 링크 추가; 본인 커밋 아님).

오늘 할 일

text cache verify 통과 확인.
본 학습 규모 결정(풀 5 epoch vs 축소), 8 GPU throughput·VRAM 실측, dataloader 병목(-c, num_workers) 확인.
평가 부분집합 확대(demo_randomized, 에피소드 수 증가).

BricksMedium

어제 한 일

BrickSim Duplo 스케일 빌드(BrickScale=2.0): expert 성공 278/300 (92.7%) vs LEGO 87.7%, job 180187(1분40초). native 빌드 10분22초.
WorkBenchMark 다단 조립 env + 순차 expert: tier1/2 80/80, tier3 29/80 (36.3%), tier4 13/80 (16.3%) — job 180209(30분20초). task_002/003은 0%.
RGB 녹화 job 180210은 make_demo_video NpzFile 재압축 해제 버그(1h43m→22s로 수정)로 TIMEOUT. 오늘 새벽 재실행: 181579(tier3/task_005), 181580(tier4/task_002) COMPLETED, wbm_eval wp0 COMPLETED(54분), v1은 PREEMPTED(13분).

오늘 할 일

tier 3–4 실패 원인 분석(task_002/003 trace): 인접 브릭 손가락 충돌 / 놓은 브릭 접촉 / joint limit 가설 검증.
2x4 180° 대칭 처리, wrist 카메라 재배치, Duplo 물리 파라미터 검증.
tier 1–2 expert demo 수집 시작 여부 결정.

SLURM 요약 (최근 24h 제출/실행)

Job ID	프로젝트(WorkDir)	JobName	실행 명령	GPU	상태	실행 시간	노드
180187	Robotics/Bricks	brk_duplo	bash scripts/eval_expert_state.sh	1	COMPLETED	00:01:40	n20
180209	Robotics/Bricks	wbm_eval	bash scripts/eval_wbm_expert.sh	1	COMPLETED	00:30:20	n43
180210	Robotics/Bricks	wbm_rgb	bash scripts/render_wbm_expert.sh	1	TIMEOUT	02:00:00	n51
181487	Robotics/LongLive/Long-WAM	lwam_rt_eval	bash scripts/robotwin2/eval_worker_b200.sh	1	COMPLETED	00:00:58	n7
181491	Robotics/LongLive/Long-WAM	lwam_rt_eval	bash scripts/robotwin2/eval_worker_b200.sh	1	COMPLETED	00:03:12	n7
181499	Robotics/LongLive/Long-WAM	lwam_rt_trsmoke	bash scripts/robotwin2/train_smoke_b200.sh	1	COMPLETED	00:11:06	n13
181500	Robotics/LongLive/Long-WAM	lwam_rt_eval	bash scripts/robotwin2/eval_worker_b200.sh	1	COMPLETED	00:20:31	n45
181534	Robotics/LongLive/Long-WAM	lwam_rt_text	bash scripts/robotwin2/precompute_text_b200.sh	2	COMPLETED	01:25:36	n7
181551	Robotics/Bricks	wbm_eval	bash scripts/eval_wbm_expert.sh	1	PREEMPTED	00:13:00	n22
181578	Robotics/Bricks	wbm_eval	bash scripts/eval_wbm_expert.sh	1	COMPLETED	00:54:08	n23
181579	Robotics/Bricks	wbm_rgb3	bash scripts/render_wbm_expert.sh	1	COMPLETED	00:16:39	n34
181580	Robotics/Bricks	wbm_rgb4	bash scripts/render_wbm_expert.sh	1	COMPLETED	00:20:25	n43
181581	Robotics/try-and-suc	gen_puck	bash scripts/gen.sh puck train 30000 22	1	CANCELLED	00:06:23	n54
181582	Robotics/try-and-suc	gen_ball	bash scripts/gen.sh ball train 30000 22	1	CANCELLED	00:05:53	n55
181604	Robotics/try-and-suc	gen_puck	bash scripts/gen.sh puck train 30000 22	1	COMPLETED	00:16:04	n22
181605	Robotics/try-and-suc	gen_ball	bash scripts/gen.sh ball train 30000 22	1	COMPLETED	00:21:10	n66
181606	Robotics/try-and-suc	tr_j_none	bash scripts/train.sh joint_none none data/puck_train data/ball_train	1	CANCELLED	00:04:40	n6
181607	Robotics/try-and-suc	tr_j_V	bash scripts/train.sh joint_V V data/puck_train data/ball_train	1	CANCELLED	00:04:10	n43
181608	Robotics/try-and-suc	tr_j_VA	bash scripts/train.sh joint_VA VA data/puck_train data/ball_train	1	CANCELLED	00:04:10	n54
181609	Robotics/try-and-suc	tr_j_VAN	bash scripts/train.sh joint_VAN VAN data/puck_train data/ball_train	1	CANCELLED	00:01:40	n19
181610	Robotics/try-and-suc	tr_j_AN	bash scripts/train.sh joint_AN AN data/puck_train data/ball_train	1	CANCELLED	00:01:10	n62
181611	Robotics/try-and-suc	tr_p_none	bash scripts/train.sh puck_none none data/puck_train	1	COMPLETED	00:08:04	n52
181612	Robotics/try-and-suc	tr_p_VA	bash scripts/train.sh puck_VA VA data/puck_train	1	CANCELLED	00:09:39	n59
181613	Robotics/try-and-suc	irp_puck	bash scripts/irp.sh puck	1	COMPLETED	00:00:53	n13
181614	Robotics/try-and-suc	irp_ball	bash scripts/irp.sh ball	1	COMPLETED	00:00:43	n52
181627	Robotics/try-and-suc	tr_j_none	bash scripts/train.sh joint_none none data/puck_train data/ball_train	1	COMPLETED	00:04:50	n53
181628	Robotics/try-and-suc	tr_j_V	bash scripts/train.sh joint_V V data/puck_train data/ball_train	1	COMPLETED	00:09:51	n73
181629	Robotics/try-and-suc	tr_j_VA	bash scripts/train.sh joint_VA VA data/puck_train data/ball_train	1	COMPLETED	00:10:28	n79
181630	Robotics/try-and-suc	tr_j_VAN	bash scripts/train.sh joint_VAN VAN data/puck_train data/ball_train	1	COMPLETED	00:10:28	n80
181631	Robotics/try-and-suc	tr_j_AN	bash scripts/train.sh joint_AN AN data/puck_train data/ball_train	1	COMPLETED	00:04:39	n6
181632	Robotics/try-and-suc	tr_p_none	bash scripts/train.sh puck_none none data/puck_train	1	COMPLETED	00:03:32	n19
181633	Robotics/try-and-suc	tr_p_VA	bash scripts/train.sh puck_VA VA data/puck_train	1	COMPLETED	00:08:47	n33
181634	Robotics/try-and-suc	eval_main	bash scripts/eval.sh main	1	COMPLETED	00:14:53	n31
181635	Robotics/try-and-suc	eval_noise	bash scripts/eval.sh noise	1	COMPLETED	00:06:48	n2
181639	Robotics/try-and-suc	eval_shuf	bash scripts/eval.sh shuffle	1	COMPLETED	00:02:04	n44
181641	Robotics/try-and-suc	eval_sweep	bash scripts/eval.sh sweep	1	COMPLETED	00:02:13	n22

Daily Progress — 2026-10-09 (금)

오늘의 우선순위

TL;DR

프로젝트별 카드

try-and-sucHigh

Long-WAM (LongLive)High

BricksMedium

활동 없음

SLURM 요약 (최근 24h 제출/실행)