---
license: mit
library_name: starvla
tags:
  - robotics
  - vision-language-action
  - florence-2
  - xtrainer2
  - retime
  - optr
---

# ReTime × xTrainer2 — Model Suite

This page indexes the **four ReTime inference bundles** for bimanual xTrainer2 tasks.
All repos are **StarVLA policy checkpoints** (consolidated DeepSpeed ZeRO-3 `.pt` files), **not** Hugging Face `transformers` models.

| # | Model | Task | Action mode | Method | Recommended checkpoint |
|---|-------|------|-------------|--------|------------------------|
| 1 | [`florence-xtrainer2-pick-bottle-uniform`](https://huggingface.co/RainyBot/florence-xtrainer2-pick-bottle-uniform) | pick bottle | **Absolute** EEF | Uniform baseline | `checkpoints/steps_35460.pt` (~60 ep) |
| 2 | [`florence-xtrainer2-pick-bottle-rel-aug-uniform`](https://huggingface.co/RainyBot/florence-xtrainer2-pick-bottle-rel-aug-uniform) | pick bottle | **Relative** EEF + image aug | Uniform + rel + aug | `checkpoints/steps_35520.pt` (~60 ep) |
| 3 | [`florence-xtrainer2-pack-phone-uniform`](https://huggingface.co/RainyBot/florence-xtrainer2-pack-phone-uniform) | pack phone | **Absolute** EEF | Uniform baseline | `checkpoints/steps_43650.pt` |
| 4 | [`florence-xtrainer2-pack-phone-residual`](https://huggingface.co/RainyBot/florence-xtrainer2-pack-phone-residual) | pack phone | **Absolute** EEF | **OPTR residual temporal sampling** (20→60 ep) | `checkpoints/steps_52380.pt` (~60 ep) |

## Shared setup

| Item | Value |
|------|-------|
| Backbone VLM | `microsoft/Florence-2-large` |
| Policy stack | StarVLA / OPTR (`QwenGR00T` + DiT-B action head) |
| Embodiment | xTrainer2, bimanual, 10 Hz |
| Cameras | `cam_high`, `cam_left_wrist`, `cam_right_wrist` |
| Robot wire | 16-D xyzw: `[xyz(3)+quat(4)+gripper(1)]` × left/right |
| Model tensor | 20-D after quaternion → rotation-6D |
| Action horizon | 16 |
| Image size | 224×224 |
| Extra deps | StarVLA tree with xTrainer2 DataConfig; `pytorch3d` |

Each repo layout:

```text
config.yaml
dataset_statistics.json
checkpoints/steps_*.pt
```

`read_mode_config` expects `config.yaml` and `dataset_statistics.json` in the **parent of `checkpoints/`**.

## How to choose a checkpoint

- **pick_bottle absolute baseline** → model 1  
- **pick_bottle relative + augmentation** → model 2  
- **pack_phone absolute baseline** → model 3  
- **pack_phone with ReTime / residual temporal sampling** → model 4  

## Serve (common)

```bash
export FLORENCE_MODEL_PATH=/path/to/Florence-2-large   # optional if using HF id
bash examples/Xtrainer2/train_files/run_xtrainer2_policy_server.sh \
  /path/to/<repo>/checkpoints/<steps_XXXXX.pt> \
  0 \
  10093
```

## Task instructions

**pick_bottle** (must match training text exactly):

- `Pick up the pink primer bottle and place it in the square dish.`
- `Pick up the pink dropper and place it in the square dish.`
- `Pick up the purple eyebrow gel tube and place it in the square dish.`
- `Pick up the silver eyebrow gel tube and place it in the square dish.`

**pack_phone**:

- `Pick up the phone, place it in the box, and close the lid.`

## License

MIT (weights). Florence-2-large follows Microsoft’s license separately.