DreamFree-LIBERO: Future Representation Distillation for Imagination-free World Action Models

DreamFree-LIBERO is a DreamFree checkpoint trained on LIBERO with a 5B video backbone.

This repository provides the full Stage 1 visual encoder and the Stage 2 action modules used by the selected LIBERO policy.

Model Details

• Architecture: WAM with a visual encoder (video DiT), readout, and action DiT.
• Video backbone: Wan2.2-TI2V-5B.
• Training data: LIBERO-Spatial, LIBERO-Object, LIBERO-Goal, and LIBERO-10.
• Inputs: two camera views, a language instruction, and robot state.
• Outputs: 32-step chunks of 7D actions.
• Encoder format: BF16 Safetensors.

Capabilities

• Imagination-free action prediction: the policy predicts actions using distilled visual representations without generating future video frames.
• Language-conditioned manipulation: the policy combines camera observations, an instruction, and proprioception to predict action chunks for LIBERO tasks.
• LIBERO evaluation: the selected recipe combines all four listed suites and uses the corresponding action normalization and image preprocessing.

How to Use

Download the checkpoint with huggingface_hub:

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="harryjo97/DreamFree-LIBERO",
    local_dir="./checkpoints/DreamFree-LIBERO",
)

The checkpoint is organized as follows:

stage1/
  config.json
  student_dit-*.safetensors
  student_dit.safetensors.index.json
stage2/
  action.safetensors
  readout.safetensors
  conditioner.safetensors
  proprio_encoder.safetensors
  config.json
  dataset_stats.json
provenance/
release.json
LICENSE.md

The Stage 1 encoder contains the full merged DiT weights; separate Stage 0 and Stage 1 adapters are unnecessary. The VAE, text encoder, and tokenizer remain upstream dependencies. Training metadata and checkpoint lineage are recorded in provenance/.

License

This DreamFree release is licensed under the MIT License. Upstream dependencies retain their respective licenses.

Citation

@article{jo2026dreamfree,
  title  = {Future Representation Distillation for Imagination-free World Action Models},
  author = {Jaehyeong Jo and Sangwon Jang and Sihyun Yu and
            Seonghyeon Ye and Yekyung Nah and Hojung Jung and
            Kimin Lee and Sung Ju Hwang},
  year   = {2026},
  journal = {arXiv preprint arXiv:2611.04527},
}