InternRobotics/InternVLA-A1-3B-RoboTwin

InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation

InternVLA-A1 integrates understanding, generation, and action experts into a unified model, which synergizes MLLMs' semantic reasoning with world-model-style dynamics prediction to guide action execution.

Building upon InternVL3 and Qwen3-VL, we instantiate InternVLA-A1 at 2B and 3B parameter scales. Covering different model scales and pre-training data configurations, we release the InternVLA-A1 series:

• InternVLA-A1-3B: pretrained on the large-scale, high-fidelity simulation data InternData-A1, together with open-source robot data (e.g. Agibot-World)
• InternVLA-A1-3B-RoboTwin: finetuned on RoboTwin 2.0 benchmark
• InternVLA-A1-3B-Pretrain-InternData-A1: pretrained on InternData-A1 only
• InternVLA-A1-2B-Pretrain-InternData-A1: pretrained on InternData-A1 only

Evaluation on RoboTwin 2.0 Simulation Benchmark

Setting: All models are jointly fine-tuned across 50 tasks (50 clean + 500 randomized demos each).

Performance Summary: InternVLA-A1-3B achieves the highest success rates across both Easy and Hard settings on the RoboTwin 2.0 Benchmark (averaged over 50 tasks).

Metric	$\pi_0$	$\pi_{0.5}$	InternVLA-A1-3B
Avg. Success (Easy)	79.98%	84.70%	88.30% 🥇
Avg. Success (Hard)	79.50%	85.02%	88.48% 🥇

🔻 Click to view detailed results for specific tasks

The table below shows success rates formatted as Easy / Hard.

Task Name	$\pi_0$	$\pi_{0.5}$	InternVLA-A1-3B
Click Bell	70.0% / 69.0%	97.0% / 93.0%	97.0% / 94.0%
Move Pillbottle Pad	83.0% / 82.0%	92.0% / 89.0%	95.0% / 99.0%
Open Laptop	90.0% / 97.0%	92.0% / 97.0%	99.0% / 99.0%
Handover Block	70.0% / 53.0%	60.0% / 59.0%	87.0% / 81.0%
Blocks Ranking Size	59.0% / 57.0%	73.0% / 77.0%	82.0% / 92.0%
Place Dual Shoes	69.0% / 76.0%	57.0% / 65.0%	93.0% / 85.0%
Stamp Seal	62.0% / 65.0%	66.0% / 73.0%	71.0% / 71.0%
Stack Bowls Three	81.0% / 75.0%	88.0% / 85.0%	86.0% / 95.0%

🔑 Key Features

Regarding model architecture, InternVLA-A1 employs a Mixture-of-Transformers (MoT) design to unifies scene understanding, visual foresight, and action execution into a single framework. It synergizes MLLM's semantic understanding with world-model-style dynamic prediction, to "imagine" the future and guide adaptive actions.

Regarding training data, We pre-train InternVLA-A1 on hybrid synthetic-real datasets spanning InternData-A1 and open-source real-world data (e.g. Agibot-World). Our hybrid synthetic-real pre-training strategy combines the scene diversity of simulation with the physical fidelity of real-world data.

Usage

Please refer to our official repo InternVLA-A1.

Demonstrations

⚡ Dynamic Manipulation

InternVLA-A1 exhibits exceptional robustness in highly dynamic scenarios.

🤖 Daily tasks

InternVLA-A1 also demonstrates superior proficiency in dexterous and fine-grained manipulation.

License and Citation

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

@article{contributors2026internvla_a1,
  title={InternVLA-A1: Unifying Understanding, Generation and Action for Robotic Manipulation},
  author={InternVLA-A1 contributors},
  journal={arXiv preprint arXiv:2601.02456},
  year={2026}
}

Acknowledgments

• Lerobot
• openpi
• InternVL
• Qwen3-VL
• COSMOS