sft_robodojo_eefabs_sana_latent_256x320_mot49500_aligned_f33fps8

Final RoboDojo ARX-X5 MoT SFT checkpoint: epoch 10, optimizer step 36,180. Original training run: sft_robodojo_mot_eefabs_sana_latent_256x320_pretrained_s49500_aligned_f33fps8.

Model and initialization

Both video and action experts were initialized by weight-only transfer from the regular MoT pretrained checkpoint epoch 9, step 49,500. SFT used the native MoT policy, with a fresh optimizer/scheduler at initial launch. The donor is logits/sana_policy_pretrained_mot_e9s49500.

• Model: SanaRWMMoTAttnResPolicy_5B_P1_D36.
• Layout: Sana-latent, with three camera views encoded separately, each at height 256 x width 320. Latent views are packed as an 8x30 strip, with semantic 2x2 spatial RoPE.
• RoPE: aligned; action_state_as_context=false.
• Three view prompts plus one action prompt; separate video/action caption embedders.
• Absolute robot-base EEF targets only, action mode ratio [0, 1, 0]; joint slots are masked.
• Column 6D rotation, fixed min=-1/max=1 normalization. XYZ and gripper use q01/q99 statistics. The included normalization applies to initial state and action targets.

Training

All 3,500 episodes; 33 source rows at 25 Hz. Configured video_fps=8 uses stride 4: rows 0,4,...,32 give 9 sampled RGB frames per view, 2 latent frames, and 32 dense action rows per window. Packed latent shape is [128, 2, 8, 30]. Training reads the precomputed full-window and freeze-padded tail caches.

8 nodes x 4 GPUs x 16 samples/GPU = global batch 512; accumulation 1. 10 epochs, AdamW lr 1e-4, cosine decay, 2,000 warmup steps, weight decay 1e-4, gradient norm clip 1, action loss weight 1.

Video/action flow shifts 5/1, standard shifts without token-count scaling. OpenWAM timestep loss weighting, min_weight=0.1.

Frozen training revision: 0928354f8505a61c1ffb21b338c4e00612d3e010, plus the source overlay identified by manifest SHA256 d93766d1c93649bf5e41bb2a2ce99cae778c9ca586cb344c513dd0b9d96d88e8. The original resize and cache behavior is preserved by this snapshot; later stretch-resize changes on the development branch were not applied to this training run.

W&B training run. Slurm job 59123 completed with exit code 0. Exact settings are archived in config.yaml and training_manifest.json.

Files

• model/pytorch_model_fsdp.bin: full native MoT state dictionary, including video and action experts.
• metadata.pth: saved epoch/step and checkpoint metadata.
• config.yaml: frozen source configuration used for SFT.
• normalization/robodojo_arx_x5_model_fps_25_f33_normalization.json: normalization artifact used for training.
• training_manifest.json: initialization, sampling, layout and training provenance.
• upload_metadata.json: file hashes and upload provenance.

This is a model-weight export; optimizer checkpoint files are not included. The archived config preserves original local paths and the training run name.