Thank you for sharing this great work!
As described in the paper, using ForeAct requires fine-tuning the VLA to accept foresight images as part of the visual input. While the pretrained foresight generator weights are available, the fine-tuned VLA weights are not, which makes it difficult to reproduce the reported results.
Would it be possible to also release the fine-tuned VLA checkpoint (e.g., the fine-tuned π₀) used in your experiments? That would be greatly appreciated!
Thank you for sharing this great work!
As described in the paper, using ForeAct requires fine-tuning the VLA to accept foresight images as part of the visual input. While the pretrained foresight generator weights are available, the fine-tuned VLA weights are not, which makes it difficult to reproduce the reported results.
Would it be possible to also release the fine-tuned VLA checkpoint (e.g., the fine-tuned π₀) used in your experiments? That would be greatly appreciated!