---
library_name: xvla
tags:
  - robotics
  - manipulation
  - bimanual
  - vision-language-action
  - maniparena
base_model: 2toINF/X-VLA-Pt
license: apache-2.0
---

# X-VLA-ManipArena v1

X-VLA soft-prompted transformer finetuned on ManipArena's `put_blocks_to_color`
task, used as the Preliminary round submission to the CVPR 2026 Embodied AI
Workshop's ManipArena competition.

## Model details

- **Base model**: `2toINF/X-VLA-Pt` (~0.9B params, Florence2 backbone + soft-prompted
  transformer)
- **Action mode**: `ee6d` (20D = 3 xyz + 6D rotation + 1 gripper per arm)
- **Training**: 20k iterations, full finetune, H200 GPU, bf16, batch=16
- **Data split**: 9:1 train/val on the official ManipArena dataset

## Open-loop eval (held-out val, 20 episodes × 3 chunks = 60 samples)

| Metric | v1 (this model) |
|---|---|
| pos_err_l_mm_mean | 46.8 |
| pos_err_r_mm_mean | 45.7 |
| rot_err_l_deg_mean | 65.8 |
| rot_err_r_deg_mean | 44.9 |
| gripper_acc | 1.00 |

## Usage

Requires the standalone [2toINF/X-VLA](https://github.com/2toINF/X-VLA) repo
cloned locally:

```python
import sys
sys.path.insert(0, "/path/to/X-VLA")

from models.modeling_xvla import XVLA
from models.processing_xvla import XVLAProcessor

model = XVLA.from_pretrained("gdgc-manip/xvla-maniparena-v1").cuda().eval()
processor = XVLAProcessor.from_pretrained("gdgc-manip/xvla-maniparena-v1")
```

Domain ID for ManipArena is **19** — set this when calling
`model.generate_actions(..., domain_id=torch.tensor([19]))`.

## Serving for ManipArena submission

See the competition adapter at:
https://github.com/... (TBD) — includes `my_policy.py` (WebSocket adapter),
`modal_serve.py` (Modal deployment), and `proxy.py` (reverse proxy for the
IP-based endpoint requirement).

Handlers that convert ManipArena's 14D RPY ↔ X-VLA's 20D 6D-rotation are in
the X-VLA fork at `X-VLA/datasets/domain_handler/maniparena.py`.

## Known limitations

- Fine-tuned on a **single task** (`put_blocks_to_color`). Other ManipArena
  tasks will use the same weights without task-specific adaptation and may
  perform poorly.
- Rotation error is ~45-66° — open-loop accumulation, not fundamentally fixed
  by more training. Closed-loop correction or shorter action horizon may help.
- Action horizon mismatch: model predicts 30 steps, ManipArena wants 50. Adapter
  pads the last 20 steps with the current proprio (hold-pose no-op).

## Team
gdgc-manip (CVPR 2026 Embodied AI Workshop, ManipArena track)