RIFT Checkpoints

Inference checkpoints for Keep the Future, Drop the Rollout: RIFT for World Action Models. Code, installation, and evaluation. This release packages existing weights for inference; checkpoint provenance is unchanged.

Dataset	Directory	Models
LIBERO	libero	FastWAM; RIFT, FastWAM-IDM and FastWAM-Joint (step 21,700); Cosmos Policy Predict2-2B
RoboCOIN, three tasks	robocoin_multitask_mm	RIFT, 10 epochs / step 9,340
Galaxea indoor cleaning	galaxea_indoor_cleaning_3cam224	FastWAM, FastWAM-Joint and RIFT, 10 epochs / step 14,830

Exact weight paths, hashes, and original source revisions are in checkpoint_manifest.json. Keep each checkpoint with its own config.yaml and dataset_stats.json. Where included, SHA256SUMS provides optional integrity checks.

Download and inference

Run from the RIFT code checkout after installing its evaluation dependencies:

hf download PoopBear/RIFT --include 'libero/rift/*' 'assets/*' --local-dir ./checkpoints
export DIFFSYNTH_MODEL_BASE_PATH=./checkpoints/assets
python experiments/libero/eval_libero_single.py \
  ckpt=./checkpoints/libero/rift/rift_step021700.pt \
  EVALUATION.dataset_stats_path=./checkpoints/libero/rift/dataset_stats.json \
  EVALUATION.task_suite_name=libero_spatial EVALUATION.task_id=0

Downloads preserve repository subdirectories: the example writes to ./checkpoints/libero/rift/. The evaluator composes the code repository's LIBERO task/data/model configs and loads the complete checkpoint; its inference defaults skip separate video/ActionDiT initialization and enable the text encoder.

Other RIFT checkpoints:

hf download PoopBear/RIFT --include 'robocoin_multitask_mm/rift_10ep/*' 'assets/*' --local-dir ./checkpoints
hf download PoopBear/RIFT --include 'galaxea_indoor_cleaning_3cam224/rift_10ep/*' 'assets/*' --local-dir ./checkpoints

Use the dataset README for the matching observation/action contract. Real-world configs retain their original structure and settings with portable paths; instantiate the packaged processor/model with the matching runtime and normalization statistics. For a complete RIFT checkpoint, use skip_dit_load_from_pretrain=true, action_dit_pretrained_path=null, and load_text_encoder=true for uncached prompts, then load all checkpoint state before inference. Relative paths resolve from your working directory. Dataset/cache paths retained in saved configs are metadata and are not downloads or prerequisites for live inference. Shared Wan VAE, text encoder, and tokenizer files are under assets/.

For a baseline, replace the include pattern with libero/fastwam/*, libero/idm/*, libero/joint/*, galaxea_indoor_cleaning_3cam224/fastwam_10ep/*, or galaxea_indoor_cleaning_3cam224/fastwam_joint_10ep/* and use its original FastWAM runtime/config. Cosmos Policy uses its upstream runtime and metadata:

hf download PoopBear/RIFT --include 'libero/cosmos-2/*' --local-dir ./checkpoints

Load PyTorch policy checkpoints with weights_only=True.

Citation

@article{zhang2026rift,
  title={Keep the Future, Drop the Rollout: RIFT for World Action Models},
  author={Zhang, Chushan and Tong, Jinguang and Li, Xuesong and Wang, Yikai and Li, Hongdong},
  journal={arXiv preprint arXiv:2608.11521},
  year={2026}
}

License

RIFT checkpoints are distributed under the Apache License 2.0 (see LICENSE). RIFT source code is distributed separately under the MIT License. See NOTICE for attribution.

The third-party Cosmos Policy files in libero/cosmos-2/ retain their original NVIDIA One-Way Noncommercial License (NSCLv1). See the included license and original model card.