RoboInter-VLM: Vision-Language Model for RoboInter Manipulation Suite

This is the flagship model of the RoboInter-VLM series, based on Qwen2.5-VL-7B-Instruct. It delivers the strongest performance among the Qwen2.5-VL variants and is the recommended default checkpoint for general use.

Developed as part of the RoboInter project. The model is fine-tuned on the RoboInter-VQA dataset for intermediate representation understanding and generation in robotic manipulation.

All Available Checkpoints

Checkpoint	Base Model	Architecture	Parameters	Description	Link
RoboInter-VLM (this repo)	Qwen2.5-VL-7B-Instruct	Qwen2.5-VL	~7B	Flagship model, recommended for best performance	https://huggingface.co/InternRobotics/RoboInter-VLM
RoboInter-VLM_qwenvl25_3b	Qwen2.5-VL-3B-Instruct	Qwen2.5-VL	~3B	Lightweight model, suitable for efficient deployment	https://huggingface.co/InternRobotics/RoboInter-VLM_qwenvl25_3b
RoboInter-VLM_llavaov_7B	LLaVA-OneVision-Qwen2-7B	LLaVA-OneVision	~7B	LLaVA-OneVision backbone with SigLIP vision encoder	https://huggingface.co/InternRobotics/RoboInter-VLM_llavaov_7B

All checkpoints are stored in safetensors format with bfloat16 precision.

Supported Tasks

These models are jointly trained on general VQA and three categories of our curated VQA tasks:

• Generation: Predicting intermediate representations such as trajectory waypoints, gripper bounding boxes, contact points/boxes, object bounding boxes (current & final), etc.
• Understanding: Multiple-choice visual reasoning about contact states, grasp poses, object grounding, trajectory selection, movement directions, etc.
• Task Planning: High-level task planning including next-step prediction, action primitive recognition, success determination, etc.

Usage

Quick Start (This Model)

from transformers import Qwen2_5_VLForConditionalGeneration, AutoProcessor

model_path = "InternRobotics/RoboInter-VLM"
model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
    model_path, torch_dtype="auto", device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_path)

For detailed usage and inference examples, please refer to the RoboInterVLM-QwenVL codebase.

LLaVA-OneVision Checkpoint

For loading and inference with the LLaVA-OneVision checkpoint, please refer to the RoboInterVLM-LLaVAOV codebase, as it requires custom model classes.

Training & Evaluation

For full training and evaluation pipelines, please refer to:

• Qwen2.5-VL models: RoboInterVLM-QwenVL
• LLaVA-OneVision model: RoboInterVLM-LLaVAOV
• VQA Dataset: RoboInter-VQA

Related Resources

• Project: RoboInter
• Annotation Data: RoboInter-Data
• VQA Dataset: RoboInter-VQA

Citation

If you find RoboInter useful in your research, please consider citing:

@article{li2026robointer,
  title={RoboInter: A Holistic Intermediate Representation Suite Towards Robotic Manipulation},
  author={Li, Hao and Wang, Ziqin and Ding, Zi-han and Yang, Shuai and Chen, Yilun and Tian, Yang and Hu, Xiaolin and Wang, Tai and Lin, Dahua and Zhao, Feng and Liu, Si and Pang, Jiangmiao},
  journal={arXiv preprint arXiv:2602.09973},
  year={2025}
}

License

Please refer to the original licenses of RoboInter, Qwen2.5-VL, and LLaVA-OneVision.