Wan2.2-5B wrist MoT — held-out val rollouts

Run mot_df_vt_a9dfilm_dforce_isolate_adaln_wrist_mixed_wan5b, checkpoint epoch=001.ckpt. Three experts (camera view, tactile L/R, wrist L/R) share one frozen Wan2.2-TI2V-5B DiT trunk with a per-expert LoRA (r=32); mixed wrist-availability training with per-clip wrist dropout p=0.2. Short rollouts predict one 4-latent window (8 predicted raw frames) from 8 GT context frames; long rollouts are 16 s (60 latent frames) keep-1 autoregressive sliding windows driven by the episode's recorded actions, from 2–3 start offsets per episode. All pixel metrics use the pad-latent-restored Wan2.2 decode.

This is epoch 1 of 30 — an early snapshot of a run that is still training, not a converged result.

All four episodes below are the run's val episodes — held out of every gradient update, but they are also the set the run selects its best checkpoint on, so these numbers are not a clean test score. The held-out test split (7 episodes, 19707 windows) is scored separately.

train 49 episodes, 162912 windows val motherboard_0510_episode_005, motherboard_0510_episode_006, motherboard_0911_episode_002_seg00, motherboard_0912_episode_003 — everything on this page test 7 episodes, 19707 windows — not shown

No-wrist val episodes

Recorded without wrist cameras (mode1_v2 session); the model runs its 3-stream path (camera view + tactile L/R).

motherboard_0510_episode_005 val motherboard_0510_episode_006 val

Short horizon — one 4-latent window (8 GT context frames → 8 predicted)

32 random windows per episode, predicted frames only (the 8 GT context frames are excluded)
stream0510_episode_0050510_episode_006
camera view22.20 dB
SSIM 0.79 · LPIPS 0.091
22.64 dB
SSIM 0.80 · LPIPS 0.082
tactile left41.47 dB
SSIM 0.96 · LPIPS 0.068
43.04 dB
SSIM 0.98 · LPIPS 0.049
tactile right43.93 dB
SSIM 0.98 · LPIPS 0.060
42.28 dB
SSIM 0.98 · LPIPS 0.082

Long horizon — 16 s autoregressive keep-1 rollout

60-latent keep-1 rollout (16 s at 15 fps), predicted frames only
stream0510_episode_005
t0=0
0510_episode_006
t0=0
camera view15.69 dB
SSIM 0.55 · LPIPS 0.404
17.14 dB
SSIM 0.60 · LPIPS 0.326
tactile left28.68 dB
SSIM 0.91 · LPIPS 0.204
28.55 dB
SSIM 0.92 · LPIPS 0.179
tactile right30.09 dB
SSIM 0.96 · LPIPS 0.163
29.91 dB
SSIM 0.95 · LPIPS 0.175

motherboard_0510_episode_005  ·  t0=0  ·  16s keep-1 @ 15.0 fps

motherboard_0510_episode_006  ·  t0=0  ·  16s keep-1 @ 15.0 fps

Short-rollout samples

motherboard_0510_episode_005

motherboard_0510_episode_006

Wrist val episodes

React v3 motherboard sessions with both wrist cameras; the model runs the full 5-stream path.

motherboard_0911_episode_002_seg00 val motherboard_0912_episode_003 val

Short horizon — one 4-latent window (8 GT context frames → 8 predicted)

32 random windows per episode, predicted frames only (the 8 GT context frames are excluded)
stream0911_episode_002_seg000912_episode_003
camera view25.02 dB
SSIM 0.85 · LPIPS 0.064
23.55 dB
SSIM 0.80 · LPIPS 0.099
wrist left15.94 dB
SSIM 0.54 · LPIPS 0.259
14.23 dB
SSIM 0.43 · LPIPS 0.346
wrist right16.57 dB
SSIM 0.48 · LPIPS 0.297
16.22 dB
SSIM 0.48 · LPIPS 0.292
tactile left45.02 dB
SSIM 0.99 · LPIPS 0.040
41.83 dB
SSIM 0.98 · LPIPS 0.058
tactile right44.76 dB
SSIM 0.99 · LPIPS 0.039
45.13 dB
SSIM 0.99 · LPIPS 0.037

Long horizon — 16 s autoregressive keep-1 rollout

60-latent keep-1 rollout (16 s at 15 fps), predicted frames only
stream0911_episode_002_seg00
t0=0
0912_episode_003
t0=0
camera view18.63 dB
SSIM 0.64 · LPIPS 0.271
17.12 dB
SSIM 0.59 · LPIPS 0.355
wrist left11.35 dB
SSIM 0.41 · LPIPS 0.560
11.04 dB
SSIM 0.44 · LPIPS 0.566
wrist right10.41 dB
SSIM 0.45 · LPIPS 0.556
10.58 dB
SSIM 0.43 · LPIPS 0.575
tactile left28.58 dB
SSIM 0.92 · LPIPS 0.206
27.35 dB
SSIM 0.89 · LPIPS 0.265
tactile right28.20 dB
SSIM 0.94 · LPIPS 0.198
27.83 dB
SSIM 0.91 · LPIPS 0.238

motherboard_0911_episode_002_seg00  ·  t0=0  ·  16s keep-1 @ 15.0 fps

motherboard_0912_episode_003  ·  t0=0  ·  16s keep-1 @ 15.0 fps

Short-rollout samples

motherboard_0911_episode_002_seg00

motherboard_0912_episode_003