DreamFAR: Real-Time and Long-Horizon Interaction with Diverse Objects
Checkpoints and sample scenes for DreamFAR, an action-conditioned world model for bimanual robot manipulation
whose dynamics are trained from scratch on robot-native teleoperation data.
| file |
what |
size |
tok256_step000200000.ckpt |
image tokenizer: a 256ร256 image โ 16ร16ร16 unit-sphere latent (weights only) |
386 MB |
dreamfar_v9_step000054000.ckpt |
dynamics model, the paper's checkpoint (EMA weights, bf16; inference only) |
4.4 GB |
spec_ee_joint_normalizer.pt |
action/state normalizer (also baked into the dynamics checkpoint) |
9 KB |
yam.urdf |
YAM bimanual station kinematics (the XDOF YAM station model), used for the end-effector actions |
14 KB |
scenes/ |
10 held-out sample start scenes to seed the model (see scenes/README_scenes.md) |
136 MB |
hf download Amazon-FAR/dreamfar --local-dir dreamfar_assets
(cd dreamfar_assets && sha256sum -c SHA256SUMS && cd scenes && sha256sum -c SHA256SUMS)
Usage (rollout, interactive demo, training): see the code repository's README. The checkpoints are PyTorch pickles
loaded by the release code (torch.load(..., weights_only=False)); only load files from this repository.
License
- Weights and sample scenes (this repository): CC BY-NC 4.0.
The tokenizer was initialized from the facebook/SSDD
F16C16_M_256
weights (CC BY-NC 4.0); the sample scenes come from the XDOF teleoperation data.
- Code: Apache-2.0, at github.com/amazon-far/dreamfar. The release code
depends on the SSDD package, which is CC BY-NC 4.0 and not included there.
yam.urdf is the XDOF YAM station model, provided for the end-effector kinematics.
Citation
@article{dreamfar2026,
title = {DreamFAR: Real-Time and Long-Horizon Interaction with Diverse Objects},
author = {Dou, Yiming and He, Botao and Loquercio, Antonio and Owens, Andrew and Malik, Jitendra and Sontakke, Sumedh Anand and Liu, Ruoshi},
journal = {arXiv preprint arXiv:2611.22001},
year = {2026}
}