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

[!NOTE] The InternVLA-A1 code has moved to the InternVLA-A1 branch.

Highlights

InternVLA-A1.5 unifies vision-language understanding, visual foresight, and action generation in one policy.

• 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.

InternVLA-A1.5 delivers strong performance across broad simulation benchmarks and real-world manipulation settings.

Model	HuggingFace	ModelScope
InternVLA-A1.5-base
InternVLA-A1.5-RoboTwin
InternVLA-A1.5-Libero
InternVLA-A1.5-DOMINO

TODO List

• [x] Release InternVLA-A1.5 training and evaluation code
• [x] Release pre-training and fine-tuning tutorials
• [x] Release simulation evaluation entries for RoboTwin, LIBERO, LIBERO-Plus, DOMINO, and SimplerEnv
• [ ] Combine InternVLA-A1, InternVLA-A1.5 and more VLAs and WAMs in the repo
• [ ] Release the example VQA data and training tutorials

Table of Contents

• Installation
• Playground
• Pre-training
• Fine-tuning
  • Fine-tuning on LeRobot V2.1 dataset
  • Fine-tuning on RoboTwin 2.0
• Evaluation & Inference

Installation

This repository has been tested on Python 3.11, CUDA 12.8, and PyTorch 2.10.0. We recommend using conda to create an isolated environment.

Please refer to the Installation Tutorial to prepare your environment, install dependencies, and patch the required HuggingFace Transformers modules for Qwen3.5 and the robot-learning policies.

Playground

Quick start with lerobot/pusht

One-line command

bash launch/internvla_a15_finetune.sh lerobot/pusht abs false

Here, abs indicates using absolute actions, and false means that the training script will use the statistics file (stats.json) provided by lerobot/pusht itself.

Pre-Training

Please refer to the Pre-training Tutorial for instructions on pretraining InternVLA-A1.5 with the InternData-A1 dataset.

The pre-training entry uses launch/internvla_a15_pretrain.sh, discovers LeRobot-format datasets under data/a1, applies configs/weight_rules_pretrain.yaml, and trains the internvla_a1_5 policy with Qwen3.5-2B, FAST action tokens, and the WAN video auxiliary branch.

Fine-tuning

Fine-tuning on LeRobot V2.1 dataset

Please refer to the LeRobot V2.1 Fine-tuning Tutorial to fine-tune InternVLA-A1.5 with real-world datasets in the LeRobot V2.1 format.

This guide walks through the complete pipeline: download dataset -> convert to LeRobot V3.0 format -> compute delta-action statistics -> fine-tune InternVLA-A1.5.

Fine-tuning on RoboTwin 2.0

Benchmark InternVLA-A1.5 on RoboTwin 2.0 with the RoboTwin 2.0 Fine-tuning Tutorial and the RoboTwin 2.0 Eval Tutorial.

Evaluation & Inference

Please refer to the evaluation guides for the complete inference and benchmark workflow:

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

For open-loop action prediction on LeRobot-format data, see tests/openloop_internvla_a1_5.py.

Real-robot inference

For real-robot deployment, we recommend using the optimized action-only inference backend. After loading the InternVLA-A1.5 config and before constructing the policy, add:

config.inference_backend, config.action_loss_only = "optimized", True

This skips WAN video-branch loading and uses the low-latency optimized backend for action prediction. Keep the standard backend only when you need future-video visualization or WAN-related inference outputs.

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 and Yang, Yuyin and Tian, Yang and Cao, Jiafei and Zhu, Hongrui and Qiu, Zherui and others},
  journal={arXiv preprint arXiv:2607.04988}
  year={2026},
}

@article{internvla_a1,
  title={InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation},
  author={Cai, Junhao and Cai, Zetao and Cao, Jiafei and Chen, Yilun and He, Zeyu and Jiang, Lei and Li, Hang and Li, Hengjie and Li, Yang and Liu, Yufei and others},
  journal={arXiv preprint arXiv:2601.02456},
  year={2026}
}

Contact

If you have any questions, feel free to submit GitHub issues or email mahaoxiang@pjlab.org.cn.

Acknowledgments

• LeRobot
• openpi
• InternVLA-A1
• Qwen3.5
• Wan