LingBot-VLA RoboTwin EEF ckpt1500

This repository contains a LingBot-VLA checkpoint for RoboTwin EEF policy evaluation.

Contents

• model-00001-of-00004.safetensors to model-00004-of-00004.safetensors: LingBot-VLA checkpoint shards.
• model.safetensors.index.json: checkpoint shard index.
• config.json: VLA model config.
• lingbotvla_cli.yaml: runtime config required by the LingBot-VLA loader.
• norm_stats/robotwin_eef.json: state/action normalization statistics.
• qwen_base/: local Qwen2.5-VL-3B-Instruct files used for config, tokenizer, processor, and base weights.

This repository does not include the RoboTwin simulator assets or the full LingBot-VLA source tree.

Usage

Download the repository:

from huggingface_hub import snapshot_download

model_dir = snapshot_download("RLinf/LingBot-VLA-RoboTwin-EEF-ckpt1500")
print(model_dir)

Start the LingBot-VLA policy server from a LingBot-VLA source checkout:

cd /path/to/VLA/code

QWEN25_PATH=<model_dir>/qwen_base \
CUDA_VISIBLE_DEVICES=0 \
OMP_NUM_THREADS=4 \
python -m deploy.lingbot_vla_policy \
  --model_path <model_dir> \
  --use_length 50 \
  --port 8008 \
  --norm_path <model_dir>/norm_stats/robotwin_eef.json

For RoboTwin hybrid evaluation, set the RoBoTwin-side VLA paths to the downloaded snapshot:

MODEL_PATH=<model_dir>
NORM_PATH=<model_dir>/norm_stats/robotwin_eef.json
QWEN25_PATH=<model_dir>/qwen_base

License Notes

The qwen_base/ directory contains Qwen2.5-VL-3B-Instruct materials. Qwen materials are distributed under the Qwen Research License. See qwen_base/LICENSE for the full license text.

This checkpoint is intended for research use.