OpenWAM Muka Franka EEF10 - pretrained fine-tune step 15000
Fine-tuned from OpenWAM/Pretrained_OpenWAM revision
e6995783e98cdba1f7be3464ff3e233592450499 on the Muka single-arm Franka
dataset.
Action contract
- Raw action/state: absolute EEF10
[xyz(3), rot6d columns(6), gripper_closedness(1)].
- Unified model action space: 80D, with active dimensions mapped to
0-9.
- The server performs quantile normalization and denormalization. Robot clients
must send and consume physical EEF10 values without applying normalization.
- Inactive unified dimensions are pinned during generation by OpenWAM PR #46.
Training and validation
- 612 episodes across 6 tasks.
- 551 task-stratified training episodes and 61 held-out validation episodes.
- 14,421 training windows and 1,489 validation windows.
- Validation uses fixed diffusion noise/timesteps and no color augmentation.
- Best held-out action loss:
0.0063270144 at step 15000.
- Code base revision:
0720891 plus the Muka LeRobot v3 reader and validation
training support used by this run.
Files
checkpoint_step_15000.safetensors: self-contained OpenWAM weights.config.yaml: model and fine-tuning configuration.normalization_stats.npy: train-split EEF10 normalization statistics.tokenizer/google/umt5-xxl/: tokenizer assets required for offline deploy.best_validation.json: held-out checkpoint selection record.
This model has passed checkpoint integrity and distributed validation checks.
Robot-side workspace and per-step safety limits remain required.
