README.md

---
license: apache-2.0
pretty_name: PRISM Data
task_categories:
  - robotics
tags:
  - humanoid
  - loco-manipulation
  - real-to-sim
  - human-object-interaction
  - counterfactual-video
  - video-to-video
  - unitree-g1
size_categories:
  - n<1K
---

# PRISM: Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation

**Zihan Wang**<sup>1,2</sup>, **Zhen Wu**<sup>1</sup>, Pieter Abbeel<sup>1,2</sup>, Rocky Duan<sup>1</sup>, Jitendra Malik<sup>1,2</sup>, Carmelo Sferrazza<sup>1</sup>, Karen Liu<sup>1,3</sup>, Guanya Shi<sup>1,4</sup>, Angjoo Kanazawa<sup>1,2</sup>

<sup>1</sup> Amazon FAR &nbsp; <sup>2</sup> UC Berkeley &nbsp; <sup>3</sup> Stanford University &nbsp; <sup>4</sup> Carnegie Mellon University

Conference on Robot Learning (CoRL 2026)

[**Project page**](https://prism-real2sim2real.github.io/) · [**Paper**](https://arxiv.org/abs/2609.38172) · [**Code**](https://github.com/amazon-far/PRISM) · [**Post**](https://x.com/Z1hanW/status/2105313719760936981)

**tl;dr:** We expand 4 real video demonstrations into diverse training data for humanoid loco-manipulation. A video-to-video model turns a recording of a person picking up a box into *counterfactual videos* of the same interaction with a bin, barrel or ball. We reconstruct these interactions in 3D, retarget them to a Unitree G1, and train in simulation. The resulting policy runs on the real robot from onboard depth only, with joystick steering, and handles diverse real objects without real-world fine-tuning.

<p align="center">
  <img src="previews/montage_2x2.gif" alt="One example per category: counterfactual video (left) and reference replay of the student training motion (right)" width="960"/>
</p>

Left: generated counterfactual video. Right: reference replay of the student training motion, time-aligned. One example per category (box, bin, barrel, ball); clip ids in `previews/preview_clips.txt`.

## Pipeline

Generated counterfactual video → 3D reconstruction → retargeted motion → learn in simulation → transfer to the real world.

Prompt template used for generation: `<VIDEO> In a real-world continuous footage. Preserve the reference video's original background, lighting and camera viewpoint. Replace the box with <CLS>, pick it up and carry with two hands.` with `<CLS>` ∈ {box, bin, barrel, ball}.

## Download

The repository contains ~7,900 small files, so a per-file download hits the Hub's API rate limit. Download the single archive instead (495 MB, sha256 `b030baccb16beb633a647366f1c1a06ae4a8b106923fa7d50d4123b68b1e6a52`):

```bash
hf download Amazon-FAR/far-prism-data far-prism-data.tar.gz --repo-type dataset --local-dir .
tar -xzf far-prism-data.tar.gz
cd far-prism-data && sha256sum --check SHA256SUMS --quiet && echo OK
```

The extracted directory has exactly the layout below. The individual files are kept in the repository for browsing.

The visual meshes inside the repository are decimated to 50k faces (bounding boxes change by < 1 mm, vertex colors kept) so that the object URDFs load everywhere without a 10 GB download. The original full-resolution reconstructions (129 meshes, ~0.5–1.2 M faces each) are in `recon-hoi-visual-meshes-original.tar.gz` (2.39 GB, sha256 `5b624d6a36786747de33c21f959b77db3c5283c6b5decb2d1271e1fb6cccefb7`); extract it over `data/recon-hoi/` to replace the decimated files.

## What is included

**129 clips** (box 35, bin 32, barrel 34, ball 28) from the 137 generated clips, keeping the ones whose teacher rollout succeeded. Per-clip index: `clips.csv`.

```text
data/train-student/
  data/motion_bank/          129 x <clip_id>.npz  (359 frames, 50 fps, robot + object world-frame trajectories, joystick commands)
  data/contact_sidecars/     per-clip hand/arm/torso contact intervals and contact points
  data/robot_assets/         G1 URDF and depth-rendering meshes
  data/checkpoints/          teacher_40000.pt (privileged teacher), box_23000.pt (student initialiser)
  vis/                       129 x <clip_id>.mp4  reference replays (1280x720, 50 fps)
data/recon-hoi/<clip_id>/
  <clip_id>_real2sim_six_axis_iou_strict_v19.obj    object visual mesh, decimated to 50k faces with vertex colors (originals: see below)
  <clip_id>_real2sim_six_axis_iou_strict_v19.urdf   object URDF (visual mesh + convex collision parts)
  meshes_collision_convex/part_*.obj                 COACD convex collision meshes (unchanged)
  retarget_qpos.npz                                  contact-anchored retarget result, qpos (217, 43) at 30 fps
data/recon-hoi/index.json                            per-clip object size, human height, SMPL scale, mesh face counts
data/raw_video/  data/train-teacher/                 under review, placeholder README only
```

Clip ids follow `prism_cf_<category>_m<seed>_v<variant>`.

## Usage

```python
root = "far-prism-data"  # extracted archive, or snapshot_download("Amazon-FAR/far-prism-data", repo_type="dataset")

import csv, numpy as np
clips = list(csv.DictReader(open(f"{root}/clips.csv")))
d = np.load(f"{root}/{clips[0]['student_motion']}")
print(d["joint_pos"].shape, d["object_pos_w"].shape, d["policy_command_xy_yaw"].shape)
```

Training code: <https://github.com/amazon-far/PRISM>.

## License

This dataset is released under the [Apache License 2.0](LICENSE).

## Citation

```bibtex
@article{wang2026counterfactual,
  title={Counterfactual Video Generation Enables Scalable Humanoid Loco-Manipulation},
  author={Wang, Zihan and Wu, Zhen and Abbeel, Pieter and Duan, Rocky and Malik, Jitendra and Sferrazza, Carmelo and Liu, C. Karen and Shi, Guanya and Kanazawa, Angjoo},
  journal={arXiv preprint arXiv:2609.38172},
  year={2026}
}
```

## Acknowledgements

We thank Chung Min Kim, Arthur Allshire, Hongsuk Choi, Isabella Yu, Junyi Zhang, Jacob Berg, Yen-Jen Wang, Sirui Chen, Charlie Cheng, Jiashun Wang, Siheng Zhao, Youjian Huang, JC Hu, Haochen Wang, Haozhi Qi and Qitao Zhao for their support and valuable feedback.