Great work! By the way, has it been compared with pi0.5?

[JXAuto123]

Great work! By the way, has it been compared with pi0.5?

[chenerg]

same question

[Zeng-Jia]

Hello, we have conducted a supplementary experiment. We compare InternVLA-A1-3B with $\pi_{0.5}$ on the RoboTwin 2.0 benchmark. Both models are jointly fine-tuned across 50 tasks (50 clean + 500 randomized demos each), The results are illustrated in the table below:

Results show our model surpasses $\pi_{0.5}$ overall. The fine-tuned checkpoint of InternVLA-A1-3B is available at https://huggingface.co/InternRobotics/InternVLA-A1-3B-RoboTwin, and we have updated this repository with the corresponding fine-tuning and evaluation code.