InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization

InternVLA-A1.5 unifies vision-language understanding, latent visual foresight, and action generation in one robot policy. It builds on a native Qwen3.5-2B VLM backbone, preserves semantic learning through VQA and subtask prediction, and attaches a lightweight unified action expert for continuous control.

This repository hosts InternVLA-A1.5-base, the base checkpoint for downstream fine-tuning and evaluation. The model uses learnable foresight tokens to query task-relevant future dynamics under the supervision of a frozen WAN2.2-5B video generation model during training. The video branch is discarded at inference, keeping deployment practical for real-time robot control.

Covering base and benchmark-specific checkpoints, we release the InternVLA-A1.5 series:

• InternVLA-A1.5-base: base checkpoint for downstream fine-tuning and evaluation
• InternVLA-A1.5-RoboTwin: fine-tuned on RoboTwin 2.0
• InternVLA-A1.5-Libero: fine-tuned on LIBERO
• InternVLA-A1.5-DOMINO: fine-tuned on DOMINO

🔑 Key Features

• 🔮 The Core: Attaches a lightweight unified action expert to a native Qwen3.5-2B VLM backbone through shared full-attention layers, while preserving modality-specific Gated DeltaNet processing.
• 🚀 The Foresight: Uses learnable foresight tokens to query task-relevant future dynamics from the shared multimodal context, supervised by a frozen WAN2.2-5B video generation model during training.
• ⚡ The Output: Discards the video branch at inference and predicts continuous action chunks through flow matching, keeping deployment latency practical.

Model Details

• Model type: Vision-Language-Action robot policy
• Backbone: Qwen/Qwen3.5-2B
• Policy type: internvla_a1_5
• Action head: unified action expert with flow-matching action generation
• Action chunk size: 50
• State/action dimension: up to 32
• Image resolution: 224 x 224
• Training auxiliary model: WAN2.2-5B video generation model for latent foresight supervision
• License: CC BY-NC-SA 4.0

Usage

Please refer to our official repo InternVLA-A-series for installation, training, fine-tuning, and evaluation.

The released training scripts use this checkpoint as the default pretrained path:

--policy.pretrained_path=InternRobotics/InternVLA-A1.5-base

For a quick LeRobot-format fine-tuning smoke test:

git clone https://github.com/InternRobotics/InternVLA-A-series.git
cd InternVLA-A-series
bash launch/internvla_a15_finetune.sh lerobot/pusht abs false

For benchmark workflows, please see:

• RoboTwin Evaluation
• LIBERO Evaluation
• LIBERO-Plus Evaluation
• DOMINO Evaluation
• SimplerEnv Evaluation

Demonstrations

InternVLA-A1.5 delivers strong performance across broad simulation benchmarks and real-world manipulation settings. The table below reports the InternVLA-A1.5 series results from the project release; benchmark-specific checkpoints should be used when reproducing fine-tuned benchmark numbers.

License and Citation

All code within this repo is released under CC BY-NC-SA 4.0. Please consider citing our project if it helps your research.

@article{internvla_a15,
  title={InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization},
  author={Ma, Haoxiang and Cai, Junhao and Xu, Xiaoxu and Li, Hao Artis, Yang, Yuyin and Tian, Yang and Cao, Jiafei and Zhu, Hongrui and Qiu, Zherui and Zhaxizhuoma and Yang, Yuqiang and Peng, Jiaqi and Wei, Xueyuan and Zhu, Yangkun and Jiang, Jiahao and Gao, Xing and Wang, Hanqing and Yuan, Feng and Li, Kailin and Zhu, Xueyue and Wang, Tai and Ding, Yan and Pang, Jiangmiao and Zeng, Jia and Zhang, Jingjing and Bowen, Zhou and Mu, Yao and Shen, Chunhua and Zhang, Weinan},
  journal={arXiv preprint arXiv:2607.04988},
  year={2026}
}