PFD LIBERO 12x12 Checkpoint

This repository contains the LIBERO 12x12 PFD checkpoint for Privileged Foresight Distillation: Zero-Cost Future Correction for World Action Models.

• Code: https://github.com/PengchengFang-cs/PFD
• Paper: https://arxiv.org/abs/2604.25859
• Checkpoint: libero_pfd_action512_partial_12x12_step62000.pt
• Model config: config.yaml
• Dataset normalization stats: dataset_stats.json

The Python package in the code release is still named fastwam for compatibility with the original training and evaluation paths.

Model

• Task: libero_uncond_2cam224_1e-4
• Model config: fastwam_pfd_action512_partial
• PFD stage: s1
• PFD training mode: action512_partial
• Partial trainable depth: action last 12 layers, video last 12 layers
• Base initialization: libero_uncond_2cam224.pt
• Training batch size: 32
• Training epochs: 30
• Selected checkpoint step: 62000

LIBERO Evaluation

Full-suite LIBERO evaluation used 50 trials per task over 40 tasks:

Suite	Successes	Success Rate
LIBERO-Spatial	493 / 500	98.60%
LIBERO-Object	496 / 500	99.20%
LIBERO-Goal	496 / 500	99.20%
LIBERO-10	477 / 500	95.40%
Overall	1962 / 2000	98.10%

The corresponding evaluation records are included under eval/.

Download

pip install -U huggingface_hub

huggingface-cli download AmberJar/PFD \
  libero_pfd_action512_partial_12x12_step62000.pt \
  config.yaml \
  dataset_stats.json \
  eval/summary.json \
  eval/task_success_rates.csv \
  --local-dir ./checkpoints/pfd_libero_12x12_step62000

Evaluation Command

From the PFD code repository:

export DIFFSYNTH_MODEL_BASE_PATH="$(pwd)/checkpoints"
export DIFFSYNTH_SKIP_DOWNLOAD=true
export LIBERO_CONFIG_PATH="$(pwd)/.libero_scratch"

python experiments/libero/run_libero_manager.py \
  task=libero_uncond_2cam224_1e-4 \
  model=fastwam_pfd_action512_partial \
  ckpt=./checkpoints/pfd_libero_12x12_step62000/libero_pfd_action512_partial_12x12_step62000.pt \
  EVALUATION.dataset_stats_path=./checkpoints/pfd_libero_12x12_step62000/dataset_stats.json \
  EVALUATION.num_trials=50 \
  MULTIRUN.num_gpus=8 \
  model.pfd.partial_unfreeze.action_last_layers=12 \
  model.pfd.partial_unfreeze.video_last_layers=12

Integrity

See SHA256SUMS and manifest.json for file hashes and provenance.

Citation

@article{fang2026pfd,
  title={Privileged Foresight Distillation: Zero-Cost Future Correction for World Action Models},
  author={Fang, Pengcheng and Chen, Hongli and Cai, Xiaohao},
  journal={arXiv preprint arXiv:2604.25859},
  year={2026}
}