LingBot-VLA overview
LingBot-VLA is a Vision-Language-Action foundation model. It receives camera observations, language instructions, and robot state, then generates robot action sequences.
Released models
| Model | Description |
|---|---|
| LingBot-VLA-4B | Pretrained checkpoint without depth distillation |
| LingBot-VLA-4B-Depth | Pretrained checkpoint with 3D geometry distilled from MoGe-2 and LingBot-Depth |
The model uses Qwen2.5-VL-3B-Instruct as the vision-language backbone and a flow-matching action expert. The released model has approximately 4B parameters.
Training and evaluation
LingBot-VLA is pretrained on 20,000 hours of real-robot interaction data across nine dual-arm robot configurations. It is evaluated with RoboTwin, LIBERO, GM-100, and other simulation and real-robot workflows. The real-robot platforms include Agibot G1, AgileX, and Galaxea R1Pro.
Next steps
- Use Installation and model download to prepare the environment and obtain a pretrained checkpoint.
- Use Prepare data and Post-training to adapt the model to a task or robot platform.
- Use Inference and deployment to run a post-trained checkpoint.
- Use Evaluation to run the open-loop, RoboTwin, LIBERO, and GM-100 workflows.