---
tags:
  - robotics
  - vision-language-action
  - forcevla
---

# ForceVLA flipbox single-view 20k deployment checkpoint

ForceVLA LoRA fine-tuned for 20,000 steps on `real_single_flexiv_tactar_flipbox`.

## Deployment files

- `params/`: final step-19999 model parameters in Orbax OCDBT format.
- `assets/real_single_flexiv_tactar_flipbox/norm_stats.json`: state, action and effort normalization statistics.
- `_CHECKPOINT_METADATA`: checkpoint metadata.
- `config/deployment_config.json`: portable model, sensor and transform contract.
- `config/forcevla_train_config_block.py`: exact `TrainConfig` entry used to build this checkpoint.

Training/deployment configuration:

- wrist image: `observation.images.left_wrist`
- external image: disabled (`image_key=None`)
- effort: `observation.effort`
- quaternion actions converted to rotation vectors
- full state enabled, effort starts at state index 8

Optimizer `train_state/` is intentionally excluded because it is not required for inference.

The norm file intentionally has top-level statistics for `state` and `actions`. Effort is not a separate top-level norm entry: the configured data transform concatenates the 6-dimensional `observation.effort` into the full state beginning at index 8.