Thank you for open-sourcing your code.

I used the checkpoints and evaluation pipeline provided in the repository directly, but my results differ substantially from those reported in the paper. I tested two tasks — bimanual_lift_ball and bimanual_handover_item_easy. The former achieves around a 40% success rate, while the latter is nearly 0%, which is both far below the paper’s numbers.

Could you please advise whether there might be any relevant configuration or setup details I could be missing?