k6_pour_full
Full fine-tuned UR5 OpenPI/Pi0.5 policy checkpoint for the k6_pour_full deployment preset.
This upload contains all available inference checkpoint steps from the local run: 10000, 20000, 30000.
The optimizer state is intentionally not uploaded by default because it is not needed for inference and is much larger than the model weights. Set UPLOAD_OPTIMIZER=1 when running the uploader if optimizer states are really needed.
Files
checkpoints/<step>/model.safetensors: PyTorch model weights for each uploaded step.checkpoints/<step>/metadata.pt: original checkpoint metadata from training.checkpoints/<step>/metadata.json: JSON-readable metadata summary for each step.checkpoints/<step>/assets/norm_stats.json: normalization stats used for inference.assets/norm_stats.json: normalization stats used for inference.configs/deployment.env: deployment preset used invla_ur5.configs/*.pyandconfigs/*.sh: source/config snapshots needed to reproduce training/inference wiring.
Training And Inference Config
- OpenPI config:
pi05_ww_k6_pour_tcp240_noforce_full_5090 - Family:
WW-k6 - Uploaded checkpoint steps:
10000,20000,30000 - Default deployment step:
30000 - Action horizon:
50 - Denoising steps used in deployment:
10 - Dataset/norm stats source:
giakhuyendihoc/ur5_pour_full_tcp240_noforce - Training time hours:
8.807756561570697
Prompt:
Pick the black and white container, place it on top of the black and red container, then pick the blue container with nuts and metal plate inside and pour the nuts and metal plate inside the black and red container until the blue container is empty
Loading
Use the local vla_ur5 / OpenPI deployment code:
TASK=k6_pour_full bash deployment/scripts/serve_policy.sh
When loading directly from this HF repo, use checkpoints/30000 as the checkpoint directory for the default final model.
The matching config name is:
pi05_ww_k6_pour_tcp240_noforce_full_5090
Notes
The config snapshots are included so other users can reconstruct the architecture and transforms even if the local repository evolves.