Awomo-0.6I

AP338 robot-pretraining weights for initializing the historical Awomo action-patch RoboTwin SFT runtime. This is a weights-only warm start, not a complete training-resume checkpoint.

• model/: 128 distributed safetensors shards; .complete: checkpoint marker.
• Action/state interface: 80D. The training adapter selects 14 valid RoboTwin slots.
• Training requires the matching historical NeMo runtime and the FLUX.2, VAE and Qwen3-VL model assets separately. The shards require its distributed Hugging Face checkpoint reader.
• Optimizer, scheduler, EMA and random-state files are not included. Disable training-state, optimizer, scheduler, data-state and EMA restoration when starting SFT.

The prepared checkpoint and RoboTwin data were exercised together in an isolated 8 × H20-3e, batch-1-per-GPU smoke test. Two optimizer steps completed and a new checkpoint was saved. This was not the original 32-GPU training configuration; downstream evaluation is not included.