sft_robodojo_mot_jointabs_f33fps16_320px_nsc

sft_robodojo_mot_jointabs_f33fps16_320px_nsc

RoboDojo ARX-X5 Sana-Video + MoT + Sana policy 2x2 multiview policy with absolute joint targets. Latest checkpoint: epoch 10, optimizer step 35000. action_state_as_context: false.

Video expert initialized from sana_v2_5b_sft_2076k_ema.pth; action expert initialized fresh. 25 FPS source, 33 source rows, 17 sampled RGB frames, 32 dense action rows, global batch 512, 10 training epochs. The config records layout and video sampling settings. AdamW lr 1e-4, cosine decay, 3,000 warmup steps, weight decay 1e-4, gradient clip 1, action loss weight 1. All 3,500 episodes are used for training.

Training run. This repository's main branch is updated at each 5,000-step milestone and at the final checkpoint.

model/pytorch_model_fsdp.bin contains the full MoT model weights. metadata.pth, config.yaml, and normalization/robodojo_arx_x5_model_fps_25_f33_normalization.json accompany them. Use this checkpoint for model loading; optimizer state is not included. upload_metadata.json records the checkpoint step and SHA256.