dish-utensil-flashwam-100ep

Dish-utensil FlashWAM (100-epoch run)

Public archival bundle for the dish-utensil FlashWAM model trained by Amin — a 100-epoch FROM-SCRATCH run (run dish_utensil_flashwam_scratch_100ep, 2026-08-19), which completed all 100 epochs (126,500 optimizer steps). Same recipe as SleepMastger/fruit-picking-flashwam-100ep (4 GPUs x batch 8 = global batch 32, cosine lr 1e-4, weight decay 1e-2, bf16, from-scratch Wan2.2 video expert + random 1-layer action expert), retargeted at dish_utensil: 95-episode train split with 5% validation holdout (40,458 frames), 1,265 steps/epoch — do not reuse other runs' step arithmetic.

Architecture: FlashWAM M1 fused-KV, fixed-RoPE, 1-layer action expert. Checkpoints were saved every 10 epochs:

• checkpoints/weights/step_012650.pt (epoch 10) — SHA-256 d324250959f66eeef80c00fb75ebd7b85921451f93fe1c8937d0af63d2ff9a1a
• checkpoints/weights/step_025300.pt (epoch 20) — SHA-256 9e1a89e939ea564fbde755dde33be8fddd34e0deb7cae466b114f53d66a7f5b2
• checkpoints/weights/step_037950.pt (epoch 30) — SHA-256 06063a3ac7ba1c14b33c53857fc43b0ae9d3aa63734c9d58ff5ecbca469ef557
• checkpoints/weights/step_050600.pt (epoch 40) — SHA-256 00c7d41918f95e6015c5478d605b7f643eef541c8e68b82be81d41d7ae15888f
• checkpoints/weights/step_063250.pt (epoch 50) — SHA-256 8f83b4a2ca5c9d0ffa2b5d15557705138083dbaecd9e5b1ee5c1eead9c07371a
• checkpoints/weights/step_075900.pt (epoch 60) — SHA-256 f8c1cae16080faae407f44531b1e975bc92d07fffe96c69830acd880b63638dd
• checkpoints/weights/step_088550.pt (epoch 70) — SHA-256 feb80010fd7a271bfc9395321739b5e397681ed65dff002244a46d2b42c7c232
• checkpoints/weights/step_101200.pt (epoch 80) — SHA-256 b36badb9e38653ee21a5f301f863f94dab4f937d7a3efde1908083331ee65088
• checkpoints/weights/step_113850.pt (epoch 90) — SHA-256 fd01140eb68c8d41df6a3c317186899464f519df26d3e74851613c50be53be44
• checkpoints/weights/step_126500.pt (epoch 100) — SHA-256 368db43199e07917ebffba4b4a98b72d638f93b769d91368e63db3b9c4a2ca09

The model has not been evaluated on a robot.

Conditioning

Exact task text (must byte-match at deployment):

pick the dishes out of the rack and place them on the table, then pick and place the spoon to the gray dish and fork to the pink dish

⚠️ This is Amin's retasked string. It differs from BOTH of George's dish repos: the SleepMastger/dish-utensil-{flashwam,fastwam}-v2 repos say "spoon in the gray dish / fork in the pink dish", and the v1 repos use "pick a utensil out of a drying rack and place it on a plate".

conditioning/text_embedding.pt is the exact cached T5 embedding consumed during training (Wan text stack, context length 128). The resolved training config sets load_text_encoder: false, so this cached tensor is part of the required inference bundle.

Input processing and normalization

• Two 256x256 RGB cameras with raw LeRobot keys image / wrist_image (NOT fruit_pick's agentview/wrist), resized to 224x224 and concatenated horizontally to 224x448.
• Dataset is LeRobot v2.1 layout (per-episode parquet + mp4), staged by training_code/stage_dish_utensil_local.sh.
• Horizon: 33 observations; 32 action transitions at 10 Hz.
• State is 8-D: eef_pos(3) + axis-angle(3) + [gripper_width/2, -gripper_width/2], width in 0..0.08 m.
• Action is 7-D: delta XYZ, delta rotation XYZ (identically zero in this data), and grasp. Grasp is {0, 1} here, NOT fruit_pick's -1/+1; min/max normalization maps either onto [-1, 1].
• dataset_stats.json contains the exact min/max normalization statistics used by this run (computed on the 95-episode train split).

Attention masks

• video_attention_mask_mode: first_frame_causal
• action_group_causal_mask_mode: group_diagonal
• Text cross-attention is enabled for the action expert.

The exact implementations, configs, split arithmetic, staging and launch scripts are included under training_code/; the resolved config is config.yaml. No license is asserted here for the bundled upstream code; its original terms continue to apply.

Base components

This weights-only checkpoint is not standalone. It references Wan-AI/Wan2.2-TI2V-5B and requires the matching Wan VAE plus the included FastWAM code/configuration. PyTorch .pt files may contain pickled objects; load only in a trusted environment.