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 windowsval motherboard_0510_episode_005, motherboard_0510_episode_006, motherboard_0911_episode_002_seg00, motherboard_0912_episode_003 — everything on this pagetest 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)
stream
0510_episode_005
0510_episode_006
camera view
22.20 dB SSIM 0.79 · LPIPS 0.091
22.64 dB SSIM 0.80 · LPIPS 0.082
tactile left
41.47 dB SSIM 0.96 · LPIPS 0.068
43.04 dB SSIM 0.98 · LPIPS 0.049
tactile right
43.93 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