LIBERO-10 Task 0: 0/50 success in inference

[purentap]

Hi, and thanks for your work. I’m running into an issue while collecting rollouts: in inference mode on LIBERO-10, Task 0, the model achieves 0/50 successes. I haven’t tested other tasks yet, but with the original OpenPI repo (same checkpoint) I see ~88% overall success. Could this indicate a bug or a configuration mismatch in this implementation?

Thanks in advance

[purentap]

Also when reviewing the episode videos, the robot behaves erratically (e.g., spinning in place, drifting off the table view, leaving the task workspace) and fails to engage with the task at all.