Error decoding tokens when inferencing on libero with pi0_fast_libero

Author: aojiaosaibanCreated Dec 19, 2025Updated Aug 24, 2026

I've read some issues about the decoding error with custom training, but i guess it's Abnormal to meet this when inferencing with official pi0_fast_libero ckpt, which has been finetuned on libero I only changed the policy model and config, here is my serve_policy.py

DEFAULT_CHECKPOINT: dict[EnvMode, Checkpoint] = {
    EnvMode.ALOHA: Checkpoint(
        config="pi05_aloha",
        dir="gs://openpi-assets/checkpoints/pi05_base",
    ),
    EnvMode.ALOHA_SIM: Checkpoint(
        config="pi0_aloha_sim",
        dir="gs://openpi-assets/checkpoints/pi0_aloha_sim",
    ),
    EnvMode.DROID: Checkpoint(
        config="pi05_droid",
        dir="gs://openpi-assets/checkpoints/pi05_droid",
    ),
    EnvMode.LIBERO: Checkpoint(
        config="pi0_fast_libero",
        dir="/data/model/pi0_fast_libero",
    ),
}

Any possible reason?

Source: Physical-Intelligence/openpi

View original on GitHub View discussion on GitHub