benchmark/tables/release_tables/aide/formulation_abs.md

Formulation ablation — absolute-coordinate read-out vs the CockpitWM formulation (EXPLANATORY, not for adoption) — AIDE test, n = 609

Generated 20260915T025941Z by aide/eval/report_formulation_abs.py from metrics_test.json / run json / pred_test.npz only (locked protocol: driver first, grid floor, HM top 60 (threshold 72.05 px), PCK @ 96 px; every run selected on val All-MPJPE of the EMA weights only, test evaluated once by the harness; CI = 10k clip-bootstrap percentile interval; group rows mean ± sd over seeds, ddof 1). Pre-registered in docs/decisions.md 2026-09-15 at 2026-09-15T02:19:20Z (before any run).

Question. The published Driver-WM Table-1 MotionBERT row (All 73.51) is worse than Driver-WM main (71.47) and worse than copying the last pose (published zero-velocity 52.89), while the same architecture trained on our keypoint protocol with the CockpitWM formulation reaches 48.46 ± 0.15 (5 seeds). Test. The same two architectures are re-trained with an ABSOLUTE-coordinate read-out — pred = out ⊙ (W, H) px, PyTorch default (non-zero) init of the read-out, no residual to the last pose, no zero-velocity start, no anchor mixture — with everything else identical: input tokens [(Δx,Δy)/200 to the last pose, x/W, y/H, conf], blocks, recipe (aide/train/run_arch_attribution.sh:24 verbatim), loss (--loss euc for the MotionBERT architecture; CockpitWM keeps all four Stage-A terms euc 1.0 + Laplace NLL 0.1 + bone 0.05 + jerk 0.01 — none depends on the residual parametrisation), seeds 42 / 123 / 456. The paired reference of every abs run is the seed-matched residual run that already exists (motionbert_kp_s<k>_v2, cwm3d_kp_s<k>_v2; untouched).

1. Test levels — published rows (labelled, typed verbatim in aide/eval/paper_rows.py), our references, and the two absolute-formulation arms

row	seed / n	All	d-n	HM	HM d-n	PCK@0.05	best val All	best ep / epochs / stop	GPU min
MotionBERT (published, Driver-WM arXiv 2605.05092v2 Table 1)	paper (single seed)	73.51	3.340	141.53	6.420	78.01	–	–	–
Driver-WM main (published)	paper (single seed)	71.47	3.240	138.03	6.270	71.66	–	–	–
Zero-Velocity (published)	paper (single seed)	52.89	2.400	139.19	6.320	85.95	–	–	–
Zero-velocity (ours, locked protocol, run zero_velocity)	analytic	52.93 [48.88, 57.25]	2.403 [2.219, 2.599]	139.23 [116.93, 167.68]	6.320 [5.308, 7.612]	85.95 [84.73, 87.15]	–	–	0
MotionBERT architecture, CockpitWM formulation (motionbert_kp, residual) — 5-seed mean ± sd	n = 5	48.46 ± 0.15	2.200 ± 0.007	130.38 ± 0.84	5.918 ± 0.038	87.14 ± 0.11	–	–	–
MotionBERT architecture, CockpitWM formulation (motionbert_kp, residual) — motionbert_kp_s42_v2	42	48.28 [44.53, 52.37]	2.192 [2.021, 2.377]	130.34 [108.20, 158.53]	5.917 [4.912, 7.196]	87.17 [86.04, 88.28]	55.550	81 / 122 / early_stop(patience=40)	4.6
MotionBERT architecture, CockpitWM formulation (motionbert_kp, residual) — motionbert_kp_s123_v2	123	48.38 [44.65, 52.46]	2.196 [2.027, 2.381]	129.16 [106.70, 157.58]	5.863 [4.844, 7.153]	87.14 [86.02, 88.27]	55.627	58 / 99 / early_stop(patience=40)	4.36
MotionBERT architecture, CockpitWM formulation (motionbert_kp, residual) — motionbert_kp_s456_v2	456	48.48 [44.74, 52.58]	2.201 [2.031, 2.387]	130.23 [107.78, 158.89]	5.912 [4.893, 7.213]	87.14 [86.00, 88.25]	55.166	90 / 131 / early_stop(patience=40)	5.17
MotionBERT architecture, ABSOLUTE formulation (motionbert_kp_abs) — motionbert_kp_abs_s42_v2	42	49.84 [45.95, 53.96]	2.262 [2.086, 2.449]	132.12 [109.41, 160.74]	5.997 [4.966, 7.297]	86.75 [85.57, 87.92]	56.240	147 / 188 / early_stop(patience=40)	4.54
MotionBERT architecture, ABSOLUTE formulation (motionbert_kp_abs) — motionbert_kp_abs_s123_v2	123	51.00 [46.66, 55.90]	2.315 [2.118, 2.538]	131.13 [108.61, 159.80]	5.952 [4.930, 7.254]	86.69 [85.45, 87.86]	56.170	152 / 193 / early_stop(patience=40)	4.79
MotionBERT architecture, ABSOLUTE formulation (motionbert_kp_abs) — motionbert_kp_abs_s456_v2	456	50.01 [46.14, 54.21]	2.270 [2.094, 2.461]	132.17 [109.29, 160.84]	6.000 [4.961, 7.301]	86.70 [85.48, 87.87]	56.304	149 / 190 / early_stop(patience=40)	4.25
MotionBERT architecture, ABSOLUTE formulation (motionbert_kp_abs) — mean ± sd	n = 3	50.28 ± 0.63	2.283 ± 0.029	131.80 ± 0.59	5.983 ± 0.027	86.71 ± 0.03	–	–	–
CockpitWM (cwm3d_kp, residual + anchors) — 5-seed mean ± sd	n = 5	48.60 ± 0.12	2.206 ± 0.005	131.57 ± 0.57	5.973 ± 0.026	87.02 ± 0.06	–	–	–
CockpitWM (cwm3d_kp, residual + anchors) — cwm3d_kp_s42_v2	42	48.77 [44.99, 52.91]	2.214 [2.042, 2.402]	132.49 [109.62, 161.16]	6.014 [4.976, 7.316]	87.00 [85.85, 88.12]	55.367	58 / 99 / early_stop(patience=40)	0.6
CockpitWM (cwm3d_kp, residual + anchors) — cwm3d_kp_s123_v2	123	48.66 [44.86, 52.81]	2.209 [2.037, 2.397]	131.05 [108.48, 159.39]	5.949 [4.925, 7.236]	86.99 [85.82, 88.14]	55.367	44 / 85 / early_stop(patience=40)	0.51
CockpitWM (cwm3d_kp, residual + anchors) — cwm3d_kp_s456_v2	456	48.46 [44.68, 52.59]	2.200 [2.028, 2.387]	131.37 [108.83, 159.91]	5.964 [4.940, 7.259]	87.13 [85.99, 88.24]	55.262	45 / 86 / early_stop(patience=40)	0.52
CockpitWM architecture, ABSOLUTE formulation (cwm3d_kp_abs) — cwm3d_kp_abs_s42_v2	42	52.76 [49.04, 56.80]	2.395 [2.226, 2.579]	131.39 [108.68, 159.94]	5.964 [4.934, 7.260]	86.13 [84.97, 87.27]	59.264	117 / 158 / early_stop(patience=40)	3.35
CockpitWM architecture, ABSOLUTE formulation (cwm3d_kp_abs) — cwm3d_kp_abs_s123_v2	123	52.02 [48.36, 56.03]	2.361 [2.195, 2.544]	129.88 [107.59, 158.22]	5.896 [4.884, 7.182]	86.47 [85.34, 87.60]	58.834	115 / 156 / early_stop(patience=40)	3.84
CockpitWM architecture, ABSOLUTE formulation (cwm3d_kp_abs) — cwm3d_kp_abs_s456_v2	456	52.73 [49.01, 56.81]	2.394 [2.225, 2.579]	134.04 [111.41, 162.77]	6.085 [5.058, 7.389]	86.19 [85.04, 87.34]	58.601	125 / 166 / early_stop(patience=40)	4.14
CockpitWM architecture, ABSOLUTE formulation (cwm3d_kp_abs) — mean ± sd	n = 3	52.50 ± 0.42	2.383 ± 0.019	131.77 ± 2.11	5.982 ± 0.096	86.26 ± 0.18	–	–	–

Published MotionBERT row vs our residual MotionBERT-kp: the table's first row and the motionbert_kp group row are the two numbers the ablation is about (73.51 vs 48.46 ± 0.15). For every row the CockpitWM formulation is the residual-to-last-pose read-out with zero init (+ anchor mixture for CockpitWM itself); the absolute formulation is the same network with the read-out replaced as described above.

Consistency check of the 5-seed reference means against arch_attribution.md §3.1: motionbert_kp recomputed 48.46 ± 0.15 vs quoted 48.46 ± 0.15 (match); cwm3d_kp recomputed 48.60 ± 0.12 vs quoted 48.60 ± 0.12 (match).

2. Paired per-clip Δ = absolute − residual (seed-matched, same architecture; errors: positive = the absolute formulation is worse; PCK: negative = worse)

MotionBERT architecture, ABSOLUTE formulation (motionbert_kp_abs) vs MotionBERT architecture, CockpitWM formulation (motionbert_kp, residual)

cell	abs s42	abs s123	abs s456	residual s42	residual s123	residual s456	Δ s42 [CI]	Δ s123 [CI]	Δ s456 [CI]	mean Δ ± sd
All-MPJPE	49.84	51.00	50.01	48.28	48.38	48.48	+1.56 [+0.91,+2.36]	+2.62 [+0.71,+5.48]	+1.53 [+0.87,+2.31]	1.90 ± 0.62
d-n	2.262	2.315	2.270	2.192	2.196	2.201	+0.071 [+0.041,+0.107]	+0.119 [+0.032,+0.249]	+0.070 [+0.039,+0.105]	0.086 ± 0.028
HM-MPJPE	132.12	131.13	132.17	130.34	129.16	130.23	+1.77 [-0.73,+4.18]	+1.97 [-0.49,+5.02]	+1.94 [-0.55,+4.74]	1.89 ± 0.10
HM d-n	5.997	5.952	6.000	5.917	5.863	5.912	+0.080 [-0.033,+0.190]	+0.089 [-0.022,+0.228]	+0.088 [-0.025,+0.215]	0.086 ± 0.005
PCK@0.05	86.75	86.69	86.70	87.17	87.14	87.14	-0.41 [-0.77,-0.13]	-0.45 [-0.93,-0.09]	-0.44 [-0.79,-0.14]	-0.44 ± 0.02
All h=1	37.01	37.57	36.28	35.04	35.01	35.17	+1.98 [+1.19,+2.85]	+2.57 [+0.81,+5.16]	+1.11 [+0.57,+1.69]	1.88 ± 0.73
All h=2	46.11	46.65	46.47	44.27	44.21	44.34	+1.84 [+0.95,+2.97]	+2.44 [+0.88,+4.83]	+2.13 [+1.19,+3.24]	2.14 ± 0.30
All h=3	53.18	53.83	53.76	51.58	51.83	51.74	+1.60 [+0.93,+2.32]	+2.00 [+0.64,+4.29]	+2.02 [+1.05,+3.24]	1.88 ± 0.24
All h=4	54.82	56.11	55.20	53.22	53.33	53.25	+1.60 [+0.61,+2.82]	+2.78 [+0.57,+5.88]	+1.95 [+1.12,+2.85]	2.11 ± 0.60
All h=5	58.06	60.85	58.35	57.30	57.54	57.91	+0.76 [+0.02,+1.53]	+3.31 [+0.28,+7.69]	+0.44 [-0.29,+1.25]	1.50 ± 1.57

CockpitWM architecture, ABSOLUTE formulation (cwm3d_kp_abs) vs CockpitWM (cwm3d_kp, residual + anchors)

cell	abs s42	abs s123	abs s456	residual s42	residual s123	residual s456	Δ s42 [CI]	Δ s123 [CI]	Δ s456 [CI]	mean Δ ± sd
All-MPJPE	52.76	52.02	52.73	48.77	48.66	48.46	+3.99 [+3.03,+4.93]	+3.36 [+2.42,+4.28]	+4.27 [+3.35,+5.17]	3.87 ± 0.46
d-n	2.395	2.361	2.394	2.214	2.209	2.200	+0.181 [+0.138,+0.224]	+0.153 [+0.110,+0.194]	+0.194 [+0.152,+0.235]	0.176 ± 0.021
HM-MPJPE	131.39	129.88	134.04	132.49	131.05	131.37	-1.10 [-5.63,+2.61]	-1.17 [-4.55,+1.96]	+2.67 [-1.28,+6.58]	0.13 ± 2.20
HM d-n	5.964	5.896	6.085	6.014	5.949	5.964	-0.050 [-0.256,+0.118]	-0.053 [-0.207,+0.089]	+0.121 [-0.058,+0.299]	0.006 ± 0.100
PCK@0.05	86.13	86.47	86.19	87.00	86.99	87.13	-0.88 [-1.30,-0.47]	-0.51 [-0.96,-0.06]	-0.93 [-1.38,-0.51]	-0.77 ± 0.23
All h=1	40.39	39.75	40.03	35.08	35.03	35.07	+5.32 [+4.40,+6.32]	+4.72 [+3.78,+5.68]	+4.96 [+4.12,+5.82]	5.00 ± 0.30
All h=2	48.96	48.41	48.81	44.84	44.66	44.50	+4.12 [+2.91,+5.25]	+3.76 [+2.66,+4.79]	+4.31 [+3.32,+5.30]	4.06 ± 0.28
All h=3	56.12	55.52	56.27	52.24	52.04	51.87	+3.88 [+2.63,+5.10]	+3.48 [+2.31,+4.64]	+4.41 [+3.25,+5.56]	3.92 ± 0.46
All h=4	57.63	56.64	57.72	53.91	53.76	53.50	+3.72 [+2.34,+5.06]	+2.88 [+1.67,+4.14]	+4.22 [+2.97,+5.46]	3.61 ± 0.67
All h=5	60.68	59.77	60.82	57.77	57.81	57.38	+2.91 [+1.54,+4.29]	+1.97 [+0.67,+3.24]	+3.44 [+2.08,+4.80]	2.77 ± 0.75

3. Pre-registered reading and outcome

Pre-stated (decisions row, before any run; applied to the seed-mean test All-MPJPE of each arm separately): (i) 60–80 px → the published MotionBERT gap is explained by the output formulation; (ii) ≤ 52.89 px (published zero-velocity) → the formulation is NOT the explanation, the difference lies elsewhere and is reported as such; (iii) anything else (52.89–60 px or > 80 px) → the formulation accounts for part of / more than the gap; descriptive, no binary conclusion. Nothing is adopted from this table; the attribution wording of arch_attribution.md is untouched; any paper sentence derived from this table needs owner review.

• MotionBERT architecture, ABSOLUTE formulation (motionbert_kp_abs) — seed-mean test All-MPJPE 50.28 ± 0.63 (n = 3/3) → (ii) EXPLANATION ELSEWHERE: seed-mean All 50.28 px is at or below the published zero-velocity 52.89: the output formulation is NOT what separates the published MotionBERT row from ours; the difference lies elsewhere (the paper's baseline pipeline / features / training details) and is reported as such. Δ to our residual re-instantiation of the same architecture: +1.83 px (seed means); Δ to the published MotionBERT row 73.51: -23.23 px; Δ to the published Driver-WM main 71.47: -21.19 px.
• CockpitWM architecture, ABSOLUTE formulation (cwm3d_kp_abs) — seed-mean test All-MPJPE 52.50 ± 0.42 (n = 3/3) → (ii) EXPLANATION ELSEWHERE: seed-mean All 52.50 px is at or below the published zero-velocity 52.89: the output formulation is NOT what separates the published MotionBERT row from ours; the difference lies elsewhere (the paper's baseline pipeline / features / training details) and is reported as such. Δ to our residual re-instantiation of the same architecture: +3.90 px (seed means); Δ to the published MotionBERT row 73.51: -21.01 px; Δ to the published Driver-WM main 71.47: -18.97 px.

Outcome sentence. Both arms stay at or below the published zero-velocity: the output formulation is NOT what separates the published MotionBERT row from ours; the explanation lies elsewhere (the paper's baseline pipeline / features / training details) and is reported as such.

4. Runs, GPU minutes, pre-registration check

run	formulation (run json config)	t_start_utc	after S0 2026-09-15T02:19:20Z?	params	GPU min
motionbert_kp_abs_s42_v2	abs	2026-09-15T02:28:54Z	yes	813,088	4.54
motionbert_kp_abs_s123_v2	abs	2026-09-15T02:33:29Z	yes	813,088	4.79
motionbert_kp_abs_s456_v2	abs	2026-09-15T02:38:20Z	yes	813,088	4.25
cwm3d_kp_abs_s42_v2	abs	2026-09-15T02:28:19Z	yes	5,771,449	3.35
cwm3d_kp_abs_s123_v2	abs	2026-09-15T02:31:43Z	yes	5,771,449	3.84
cwm3d_kp_abs_s456_v2	abs	2026-09-15T02:35:37Z	yes	5,771,449	4.14
total (new runs)					24.91

Provenance

• New run dirs only: aide/results/runs/motionbert_kp_abs_s<k>_v2, aide/results/runs/cwm3d_kp_abs_s<k>_v2 (launcher aide/train/run_formulation_abs.sh, gate before every launch: free VRAM ≥ 3 + 4 GB, MemAvailable ≥ 28 GB, < 2 trainers alive on the host incl. the BATON CAN-arm trainers, OMP/MKL 4; pids in aide/results/formulation_abs/pids.txt). The comparison runs motionbert_kp_s*_v2, cwm3d_kp_s*_v2, zero_velocity are the existing ones (not retrained).
• Code: --formulation {residual,abs} in aide/train/train_kp.py; heads.AbsoluteReadout, HorizonHead(zero_init=), CWM3D(formulation=), MotionBERTKP(formulation=). The default residual path is asserted byte-identical to the pre-change code by aide/tests/test_formulation_abs.py (fixture aide/tests/fixtures/formulation_abs_prechange.pt generated from the unmodified files before the edit).
• Published numbers: aide/eval/paper_rows.py (typed verbatim from arXiv 2605.05092v2 Table 1; single-seed rows). No AIDE test clip entered any training or selection; every abs run was selected on val All-MPJPE only and its test split evaluated once.