---
license: apache-2.0
library_name: transformers
tags:
- robotics
- vla
- lingbot-vla
- lerobot
- manipulation
---

# LingBot-VLA 2.0 — Stack-the-cubes finetune

Finetune of **LingBot-VLA 2.0** (6B, MoE action expert) on the mixed
[LGG100/Stack-the-cubes](https://huggingface.co/datasets/LGG100/Stack-the-cubes) +
[LGG100/Stack-the-cubes-v2](https://huggingface.co/datasets/LGG100/Stack-the-cubes-v2)
datasets (Unitree G1 Dex1, dual-arm cube stacking).

Code & real-robot serving: https://github.com/YCC-DAVID/robotic

## Checkpoints

Each `global_step_*/` folder holds the deployable HF-format weights
(`model-0000x-of-00006.safetensors` + tokenizer/config), loadable directly.

| Folder | Step |
|---|---|
| `global_step_5000/`  | 5000  |
| `global_step_10000/` | 10000 |
| `global_step_15000/` | 15000 |
| `global_step_20000/` | 20000 (final) |

## Training

- Base: LingBot-VLA 2.0 (fp32), 16-dim state/action (14 arm joints + 2 grippers)
- 2xA100-80GB, FSDP2, Muon optimizer, `L1_fm` loss, `bounds_99_woclip` norm, absolute actions
- 20000 steps, lr 1e-4, with depth + DINO-video distillation and future-image prediction
- Final loss ~0.11 (VLA_Loss)

Only the deployable `hf_ckpt` weights are published (optimizer states omitted).