Daily Progress — 2026-10-02 (금)

프로젝트별 카드

Zero-WAM High

어제 한 일

SLURM 9건 (COMPLETED 5 / PREEMPTED 3 / CANCELLED 1): 8 GPU 규모 ICL 평가 165391 (7 GPU, 1d 01:20:03, 완료), lang 평가 태스크별 1 GPU job(open_microwave 167170 13:45:47 완료, stack_blocks_three 167124 own 09:44:39 완료). extra 평가 5회 preempt(손실 에피소드 1개씩, START_SEED 이어받기).
50K 최종 표 (seed 0, n=100/태스크): ICL macro 27.00→22.57 (p≈0.055), lang 20.67→17.57 (p≈0.13), Δdemo +6.33→+5.00. stamp_seal 47→45, move_stapler_pad 40→36, place_empty_cup 28→21, open_microwave 21→16, stack_blocks_three 0→0.
해석: 공식 50K를 채울수록 논문(36.36)과 격차 −9.29→−13.79pp. 후보 = 과적합(샘플당 반복 119→186회), Adam 상태 재시작 2회, 해상도/서브샘플/시드.
문서: 260929/analysis-zerowam_experiments_synthesis.md, exp-zerowam_wopretrain_50k.md, context.md 갱신 (전 실험 종합: 헤드라인 격차 +29.50→+6.24pp, ZW 승 태스크 0개).
train.py에 num_steps가 save_interval 배수가 아닐 때 최종 스텝 저장 추가.

오늘 할 일

36K/40K/44K/48K ckpt ICL 평가 → 32K→36K 단조 감소면 과적합, 36K에서만 꺾이면 Adam 재시작.
영상 교체 + 캡션 고정 대조(캡션/영상 교란 분리), video 마지막 프레임만 조건 (iii).
긴 평가는 own으로, 한 job에 여러 태스크 묶지 않기 (165391에서 최대 6 GPU 유휴 발생).

EmbodiedSWE High

어제 한 일

SLURM 62건 (COMPLETED 60 / PREEMPTED 2): one_leg seed 0–10 검증, 물리 설정 벤치 6조합(bench_A~F), fast/fast2/fast3 해법 스텝 축소 반복, square_table sq/sq2 seed 0–5 및 녹화. 전부 1 GPU extra, 1h 이내.
one_leg 속도 개선: 작업당 14~16분→5.5~7.1분. 해법 실행 790초(A: SDF512/iter192)→348초(C: SDF512/iter64). 비용은 솔버 반복이 지배(SDF 해상도 영향 없음); iter 32(522초)·dt 1/120은 조이기 실패. v3(손가락 상태 대기·10°×9 회전) 성공 8/8, 공식 채점 success.
square_table(다리 4개): scene/grader를 다리 N개로 일반화, 환경 assembly.furniture_square_table.* 등록. v1 1/3 (옆 다리 간섭으로 집기 실패) → v2(먼 구멍 우선, 손가락 간격 50mm, 3회 재시도) 5/6, 다리 23/24, 에피소드 17~28분.
문서: 261001/exp-furniture_one_leg_speedup.md, exp-furniture_square_table.md, context.md.

오늘 할 일

첫 조이기 사이클 기울어짐 보정(회전 속도 감소 또는 재세우기), 회전량 90°→120~150° 검토로 추가 단축.
desk / stool 이식. IKEA 테이블 4다리 solve.py 없음. Q_STAND 하드코딩 → med/high 랜덤화 일반화 미해결.

GPT-Policy High

어제 한 일

SLURM 16건 (COMPLETED 12 / FAILED 4): 300s 예산 seed 0–4 두 라운드 (167125–129, 167548–553, 167942–946). 167549–552 4건 FAILED는 Codex 사용량 한도(usageLimitExceeded, 재개 10/05)로 중단 → 평가 무효.
settle 수정(쥔 상태 그리퍼 떨림 무시): settle 실패 13–36회→0–2회, 그러나 300s 성공률 1/5로 불변 → 병목이 시간에서 물리 이상(yaw 불일치, 다리 튕김)으로 이동.
그리퍼 FurnitureBench 기본값(손가락당 13N, friction 15, stiffness 2000) 적용 후 새 Codex 계정 재실행: 2/5 성공 (seed 0: 55 dec/198s/$7.7, seed 4: 49 dec/209s/$6.7). yaw 불일치 3건(87°/36°/180°) 소멸 → 미끄러운 그립이 원인.
잔여 실패: 얕게 꽂은 다리를 놓으면 넘어짐(seed 2, 3), 탁판 뒤집힘 1건. API 환산 $7–25/episode.
분석: EmbodiedSWE one_leg 11/11 vs 본 프로젝트 2/5는 정답 pose 관측 + 오프라인 디버깅 프로그램 + 관대한 판정(나사축만) vs RGB 온라인 VLM + 원판정이라 비교 불가. 빌릴 것: ~12mm 눌러 넣고 놓기, 몸통 아래쪽 재파지, 10° 증분 회전.

오늘 할 일

지시문/screw 도구에 삽입 깊이·재파지 위치 반영, 측면 파지 첫 나사산 체결 모드 추가 → seed 0–4 재실행(n 확대 전 안정화).
기존 run을 EmbodiedSWE식 판정으로 재채점. 미해결: give_up 후 return_home IK 실패 → job FAILED, 스레드 refresh 턴이 비용 44%.

ICL_CVPR Medium

어제 한 일

Harness + ICL-WAM 포지셔닝 문헌 검증: 로컬 HTML 124건 + awesome-robots-icl README 대조, 후보 약 45편 원문 확인 (본문 12편, 초록 다수, 웹 검색 2회). SLURM/commit 없음.
틀린 주장: "executor 채널 언어/2D 포인트뿐" — π0.7(subgoal 이미지), MaP-WAM(visual plan; RMBench 83.3 vs 77.1), PonderPounce, WorldScape 2.0, Skild S1(사람 영상 1개, unseen 66% vs 언어 VLA 9%), GEN-1.5, ReCAP(retrieval). 검증기는 FaRe(RoboTwin Hard 90.2→92.4, 실로봇 71.7→83.3%), CheckVLA 선행.
남는 차별점: "agent가 데모 영상 구간을 선택·재선택해 ICL-WAM context로 전달" 조합 자체는 발견 못 함.
최대 리스크: executor가 영상을 따라가는지 미검증 (Zero-WAM 영상 기여 +7.5pt; 위 Zero-WAM 교체 대조는 −13pp로 일부 지지). agent가 자른 clip은 학습 분포 OOD.

오늘 할 일

hero 실험 선정 (stack three 9.0% 등 long-horizon vs 접촉 조립 16.7%), baseline 확정 (monolithic / ReCAP / MaP-WAM / HarnessWAM), Zero-WAM 공개 ckpt 확인.

그 외 프로젝트

scene_bench (scene-mem-gen 포함) · scene_bench_paper · FurnitureBench · R2S · lingbot-va — No activity (커밋·claude/ 문서·SLURM 없음).

SLURM 요약 (87건)

프로젝트별: EmbodiedSWE 62 (COMPLETED 60, PREEMPTED 2) · GPT-Policy 16 (COMPLETED 12, FAILED 4) · Zero-WAM 9 (COMPLETED 5, PREEMPTED 3, CANCELLED 1). 현재 RUNNING/PENDING 없음. 프로젝트명은 WorkDir 기준, 실행 명령은 SubmitLine(template.sh 이후), JobName 병기.

Job ID	프로젝트	JobName	실행 명령	GPU	상태	실행 시간	노드
165391	Zero-WAM	zw50k_eval_icl	bash script/eval_zerowam_robotwin.sh ckpt:checkpoint_step_50000 icl	7	COMPLETED	1-01:20:03	n23
166677	Zero-WAM	zw50k_lres_open_microwave	START_SEED=100007 EPISODE_OFFSET=6 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 open_microwave lang 94 0	1	PREEMPTED	12:57:30	n48
166679	Zero-WAM	zw50k_lres_place_bread_basket	START_SEED=100021 EPISODE_OFFSET=16 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 place_bread_basket lang 84 0	1	COMPLETED	09:29:44	n44
166680	Zero-WAM	zw50k_lres_place_empty_cup	START_SEED=100020 EPISODE_OFFSET=18 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 place_empty_cup lang 82 0	1	COMPLETED	07:58:19	n55
166681	Zero-WAM	zw50k_lres_stack_blocks_three	START_SEED=100007 EPISODE_OFFSET=7 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 stack_blocks_three lang 93 0	1	PREEMPTED	05:50:30	n58
166996	EmbodiedSWE	one_leg_s5	bash tmp/one_leg_seed.sh 5	1	COMPLETED	00:27:35	n46
166997	EmbodiedSWE	one_leg_s3	bash tmp/one_leg_seed.sh 3 record	1	COMPLETED	00:19:06	n49
167006	EmbodiedSWE	one_leg_s5	bash tmp/one_leg_seed.sh 5	1	COMPLETED	00:15:17	n66
167007	EmbodiedSWE	one_leg_s6	bash tmp/one_leg_seed.sh 6	1	COMPLETED	00:15:34	n67
167008	EmbodiedSWE	one_leg_s7	bash tmp/one_leg_seed.sh 7	1	COMPLETED	00:14:54	n46
167009	EmbodiedSWE	one_leg_s8	bash tmp/one_leg_seed.sh 8	1	COMPLETED	00:15:22	n49
167010	EmbodiedSWE	one_leg_s9	bash tmp/one_leg_seed.sh 9	1	COMPLETED	00:16:08	n51
167011	EmbodiedSWE	one_leg_s10	bash tmp/one_leg_seed.sh 10	1	COMPLETED	00:15:14	n52
167014	Zero-WAM	zw50k_lres2_stack_blocks_three	START_SEED=100040 EPISODE_OFFSET=35 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 stack_blocks_three lang 65 0	1	PREEMPTED	03:42:00	n40
167024	EmbodiedSWE	one_leg_verify	bash tmp/one_leg_verify.sh	1	COMPLETED	00:00:06	n52
167025	EmbodiedSWE	one_leg_s0	bash tmp/one_leg_seed.sh 0	1	COMPLETED	00:14:53	n47
167026	EmbodiedSWE	one_leg_s1	bash tmp/one_leg_seed.sh 1	1	COMPLETED	00:14:54	n51
167027	EmbodiedSWE	one_leg_s2	bash tmp/one_leg_seed.sh 2	1	COMPLETED	00:15:26	n54
167028	EmbodiedSWE	one_leg_s3	bash tmp/one_leg_seed.sh 3	1	COMPLETED	00:14:39	n57
167029	EmbodiedSWE	one_leg_s4	bash tmp/one_leg_seed.sh 4	1	COMPLETED	00:15:10	n58
167038	EmbodiedSWE	one_leg_verify	bash tmp/one_leg_verify.sh	1	COMPLETED	00:00:07	n54
167043	EmbodiedSWE	one_leg_verify	bash tmp/one_leg_verify.sh	1	COMPLETED	00:15:28	n54
167111	Zero-WAM	zw50k_lres3_stack_blocks_three	START_SEED=100060 EPISODE_OFFSET=52 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 stack_blocks_three lang 48 0	1	CANCELLED	01:08:23	n37
167124	Zero-WAM	zw50k_lres4_stack_blocks_three	START_SEED=100064 EPISODE_OFFSET=56 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 stack_blocks_three lang 44 0	1	COMPLETED	09:44:39	n7
167125	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:12:51	n15
167126	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:20:52	n43
167127	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:14:04	n46
167128	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:23:02	n47
167129	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:18:40	n49
167170	Zero-WAM	zw50k_lres2_open_microwave	START_SEED=100076 EPISODE_OFFSET=49 bash script/eval_zerowam_robotwin_task.sh ckpt:checkpoint_step_50000 open_microwave lang 51 0	1	COMPLETED	13:45:47	n51
167548	GPT-Policy	gpt-policy-furniture-smoke	scripts/smoke_furniture.sbatch	1	COMPLETED	00:00:33	n17
167549	GPT-Policy	gpt-policy-furniture	--time=01:00:00 --dependency=afterok:167548 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	FAILED	00:19:35	n46
167550	GPT-Policy	gpt-policy-furniture	--time=01:00:00 --dependency=afterok:167548 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	FAILED	00:19:16	n47
167551	GPT-Policy	gpt-policy-furniture	--time=01:00:00 --dependency=afterok:167548 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	FAILED	00:19:05	n48
167552	GPT-Policy	gpt-policy-furniture	--time=01:00:00 --dependency=afterok:167548 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	FAILED	00:18:35	n52
167553	GPT-Policy	gpt-policy-furniture	--time=01:00:00 --dependency=afterok:167548 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:13:21	n53
167554	EmbodiedSWE	bench_A_512_192_240	bash tmp/one_leg_bench.sh A_512_192_240 0 --sdf-res 512 --iters 192	1	COMPLETED	00:00:01	n72
167555	EmbodiedSWE	bench_B_256_192_240	bash tmp/one_leg_bench.sh B_256_192_240 0 --sdf-res 256 --iters 192	1	COMPLETED	00:00:01	n73
167556	EmbodiedSWE	bench_C_512_64_240	bash tmp/one_leg_bench.sh C_512_64_240 0 --sdf-res 512 --iters 64	1	COMPLETED	00:00:00	n66
167557	EmbodiedSWE	bench_D_256_64_240	bash tmp/one_leg_bench.sh D_256_64_240 0 --sdf-res 256 --iters 64	1	COMPLETED	00:00:00	n72
167558	EmbodiedSWE	bench_E_256_32_240	bash tmp/one_leg_bench.sh E_256_32_240 0 --sdf-res 256 --iters 32	1	COMPLETED	00:00:00	n74
167559	EmbodiedSWE	bench_F_256_64_120	bash tmp/one_leg_bench.sh F_256_64_120 0 --sdf-res 256 --iters 64 --dt 0.0083333333	1	COMPLETED	00:00:00	n75
167567	EmbodiedSWE	bench_A_512_192_240	bash tmp/one_leg_bench.sh A_512_192_240 0 --sdf-res 512 --iters 192	1	COMPLETED	00:13:47	n66
167568	EmbodiedSWE	bench_B_256_192_240	bash tmp/one_leg_bench.sh B_256_192_240 0 --sdf-res 256 --iters 192	1	COMPLETED	00:14:47	n68
167569	EmbodiedSWE	bench_C_512_64_240	bash tmp/one_leg_bench.sh C_512_64_240 0 --sdf-res 512 --iters 64	1	COMPLETED	00:06:32	n72
167570	EmbodiedSWE	bench_D_256_64_240	bash tmp/one_leg_bench.sh D_256_64_240 0 --sdf-res 256 --iters 64	1	COMPLETED	00:06:07	n74
167571	EmbodiedSWE	bench_E_256_32_240	bash tmp/one_leg_bench.sh E_256_32_240 0 --sdf-res 256 --iters 32	1	COMPLETED	00:09:23	n75
167572	EmbodiedSWE	bench_F_256_64_120	bash tmp/one_leg_bench.sh F_256_64_120 0 --sdf-res 256 --iters 64 --dt 0.0083333333	1	COMPLETED	00:06:30	n58
167588	EmbodiedSWE	fast_s0	bash tmp/one_leg_fast.sh it64 0 --iters 64	1	COMPLETED	00:05:43	n80
167589	EmbodiedSWE	fast_s1	bash tmp/one_leg_fast.sh it64 1 --iters 64	1	COMPLETED	00:05:10	n81
167590	EmbodiedSWE	fast_s2	bash tmp/one_leg_fast.sh it64 2 --iters 64	1	COMPLETED	00:07:00	n82
167591	EmbodiedSWE	fast_s3	bash tmp/one_leg_fast.sh it64 3 --iters 64	1	COMPLETED	00:05:39	n77
167592	EmbodiedSWE	fast_s4	bash tmp/one_leg_fast.sh it64 4 --iters 64	1	COMPLETED	00:05:22	n75
167593	EmbodiedSWE	fast_s5	bash tmp/one_leg_fast.sh it64 5 --iters 64	1	COMPLETED	00:05:08	n78
167608	EmbodiedSWE	fast2_s0	bash tmp/one_leg_fast2.sh v2it64 0 --iters 64	1	COMPLETED	00:07:05	n68
167609	EmbodiedSWE	fast2_s1	bash tmp/one_leg_fast2.sh v2it64 1 --iters 64	1	COMPLETED	00:06:31	n70
167610	EmbodiedSWE	fast2_s2	bash tmp/one_leg_fast2.sh v2it64 2 --iters 64	1	COMPLETED	00:13:25	n66
167611	EmbodiedSWE	fast2_s3	bash tmp/one_leg_fast2.sh v2it64 3 --iters 64	1	COMPLETED	00:07:09	n75
167612	EmbodiedSWE	fast2_s4	bash tmp/one_leg_fast2.sh v2it64 4 --iters 64	1	COMPLETED	00:06:20	n78
167613	EmbodiedSWE	fast2_s5	bash tmp/one_leg_fast2.sh v2it64 5 --iters 64	1	COMPLETED	00:06:05	n82
167619	EmbodiedSWE	fast3_s0	bash tmp/one_leg_fast3.sh v3it64 0 --iters 64	1	COMPLETED	00:07:06	n68
167620	EmbodiedSWE	fast3_s1	bash tmp/one_leg_fast3.sh v3it64 1 --iters 64	1	COMPLETED	00:06:35	n69
167621	EmbodiedSWE	fast3_s2	bash tmp/one_leg_fast3.sh v3it64 2 --iters 64	1	COMPLETED	00:05:30	n71
167622	EmbodiedSWE	fast3_s3	bash tmp/one_leg_fast3.sh v3it64 3 --iters 64	1	COMPLETED	00:06:50	n75
167623	EmbodiedSWE	fast3_s4	bash tmp/one_leg_fast3.sh v3it64 4 --iters 64	1	COMPLETED	00:05:58	n76
167624	EmbodiedSWE	fast3_s5	bash tmp/one_leg_fast3.sh v3it64 5 --iters 64	1	COMPLETED	00:06:24	n78
167625	EmbodiedSWE	fast3_s6	bash tmp/one_leg_fast3.sh v3it64 6 --iters 64	1	COMPLETED	00:06:43	n66
167626	EmbodiedSWE	fast3_s7	bash tmp/one_leg_fast3.sh v3it64 7 --iters 64	1	COMPLETED	00:06:50	n70
167645	EmbodiedSWE	one_leg_verify	bash tmp/one_leg_verify.sh	1	COMPLETED	00:05:54	n47
167942	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:11:26	n9
167943	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:14:27	n15
167944	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:25:02	n25
167945	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:20:23	n23
167946	GPT-Policy	gpt-policy-furniture	--time=01:00:00 scripts/run_furniture.sbatch --input-json configs/tasks/furniture-one-leg-screw-tool.json	1	COMPLETED	00:10:45	n33
168167	EmbodiedSWE	one_leg_rec	bash tmp/one_leg_seed.sh 0 record	1	COMPLETED	00:09:42	n78
168341	EmbodiedSWE	sq_s0	bash tmp/sq_run.sh v1 0 record	1	PREEMPTED	00:10:30	n74
168342	EmbodiedSWE	sq_s1	bash tmp/sq_run.sh v1 1	1	COMPLETED	00:20:18	n80
168343	EmbodiedSWE	sq_s2	bash tmp/sq_run.sh v1 2	1	COMPLETED	00:18:20	n81
168344	EmbodiedSWE	sq_s3	bash tmp/sq_run.sh v1 3	1	COMPLETED	00:16:33	n82
168368	EmbodiedSWE	sq_s3_rec	bash tmp/sq_run.sh v1trace 3 record	1	PREEMPTED	00:13:00	n39
168384	EmbodiedSWE	sq_s3_rec	bash tmp/sq_run.sh v1trace 3 record	1	COMPLETED	00:21:05	n26
168417	EmbodiedSWE	sq2_s0	bash tmp/sq_run2.sh v2 0 record	1	COMPLETED	00:28:20	n11
168418	EmbodiedSWE	sq2_s1	bash tmp/sq_run2.sh v2 1	1	COMPLETED	00:23:43	n17
168419	EmbodiedSWE	sq2_s2	bash tmp/sq_run2.sh v2 2	1	COMPLETED	00:21:30	n26
168420	EmbodiedSWE	sq2_s3	bash tmp/sq_run2.sh v2 3	1	COMPLETED	00:21:47	n38
168421	EmbodiedSWE	sq2_s4	bash tmp/sq_run2.sh v2 4	1	COMPLETED	00:17:58	n47
168422	EmbodiedSWE	sq2_s5	bash tmp/sq_run2.sh v2 5	1	COMPLETED	00:19:38	n59

Daily Progress — 2026-10-02 (금)

오늘의 우선순위

TL;DR

프로젝트별 카드

Zero-WAM High

어제 한 일

오늘 할 일

EmbodiedSWE High

어제 한 일

오늘 할 일

GPT-Policy High

어제 한 일

오늘 할 일

ICL_CVPR Medium

어제 한 일

오늘 할 일

그 외 프로젝트

SLURM 요약 (87건)