Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience

Ropedia W0 is an egocentric world foundation model that learns loco-manipulation from 10k hours of human experience and transfers it across embodiments, from robot arms to humanoids and from grippers to dexterous hands, through one shared kinematic interface of camera poses and projected skeletons.

• Model: One World Model, Multiple Embodiments. A unified egocentric loco-manipulation world foundation model spanning humans and diverse robots, from robotic arms to humanoids with grippers or dexterous hands, conditioned on camera poses and projected kinematic skeletons.
• Data: 13K Hours of Unified Experience. 10K hours of human and 3K hours of robot egocentric video, with metric camera trajectories, kinematic skeletons, and hierarchical captions.
• Scaling: Scaling Human Experience to Robots. More human pre-training data consistently improves world prediction for both humans and robots.
• Performance: State of the Art Across Embodiments. Leading results on five robot benchmarks and held-out human videos, spanning zero-shot, few-shot, and full-data post-training settings.
• Applications: From Prediction to Interaction and Planning. Supports real-time, minute-long interactive VR rollouts through distillation, alongside policy evaluation, model-based planning, and world-action modeling.

• Release technical report
• Release inference code (5B)
• Release 5B checkpoints, pre-trained and mid-trained (Hugging Face)
• Release Xp-Teleop teleoperation data (Hugging Face)
• Release A14B checkpoints and inference code
• Release benchmark fine-tuned checkpoints (5B and A14B on Xp-Teleop, ω-HOME, Ψ0 and GR-1)
• Release evaluation code and benchmark data
• Release post-training code

• System: Linux.
• Hardware: an NVIDIA GPU with at least 40 GB of memory for the 5B model (31 GB peak at 640×480, 81 frames). The code has been verified on H200.
• Software:
  • Conda with Python 3.12.
  • An NVIDIA driver supporting CUDA 12.8. For CUDA 13 drivers see ENVIRONMENT.md.
  • No custom CUDA extensions are needed; attention uses PyTorch SDPA.

1. Clone the repo:

   git clone https://github.com/Ropedia/ropedia-w0.git
   cd ropedia-w0

2. Create the environment and install the dependencies (torch 2.11 with CUDA 12.8):

   conda create -n ropedia-w0 python=3.12 -y
   conda activate ropedia-w0
   pip install torch==2.11.0 torchvision==0.26.0 --index-url https://download.pytorch.org/whl/cu128
   pip install -r requirements.txt

3. Download the Wan2.2-TI2V-5B base model (the full repository):

   hf download Wan-AI/Wan2.2-TI2V-5B --local-dir models/Wan2.2-TI2V-5B

   Any other location works with --base_model_dir (or BASE_MODEL_DIR for the scripts).

The Ropedia W0 checkpoints are hosted on Hugging Face at ropedia-ai/Ropedia-W0. Place them under weights/ (or pass their paths to the scripts below).

Run inference on one clip with the provided script. examples/ ships 18 ready-to-run clips from Xp-Teleop, ω-HOME, AgiBot World, GR-1 and human egocentric video (Xperience-10M, EgoDex, OpenAoE):

CFG=3.0 FPS=30 bash scripts/run_single_case_5b.sh examples/xp_teleop_playdoh_tub weights/w0-5b-midtrain.safetensors outputs/playdoh_tub
# robot clips: CFG=3.0 FPS=30 (GR-1 clips also NUM_FRAMES=49 FPS=10); human clips: the defaults (cfg 5.0, fps 16)

A clip directory holds video.mp4 (its first frame conditions the model), data.npz (camera intrinsics and camera-to-world poses), skeleton.mp4 (the rendered skeleton condition) and prompt.txt. To render skeleton.mp4 from human hand keypoints, run python tools/render_skeleton.py [CLIP_DIR].

The script runs inference_5b.py with the following arguments:

python inference_5b.py \
    --video [CLIP_DIR]/video.mp4 \
    --camera_params [CLIP_DIR]/data.npz \
    --skeleton_video [CLIP_DIR]/skeleton.mp4 \
    --prompt [TEXT_PROMPT] \
    --ckpt [CKPT] \
    --num_frames 81 --height 480 --width 640 --fps [FPS] \
    --cfg_scale [CFG] --num_inference_steps 50 --seed 0 \
    --output_dir [OUTPUT_DIR]

🎥 Conditions

• --camera_params: data.npz with per-frame intrinsics and camera-to-world poses (OpenCV convention); poses are taken relative to the first frame of the window.
• --skeleton_video: skeleton condition video on a black background, frame-aligned with the window.
• --scene_scale: multiplier on the camera translation (default 1.0).
• --disable_camera / --disable_skeleton: run without one of the conditions.

🎬 Window and sampling

• --num_frames, --height, --width: 81 frames at 640×480 for the released checkpoints.
• --start_frame: first frame of the window in the source video (default 0).
• --fps: frame rate of the written videos (16 for human egocentric data, 30 for robot data).
• --cfg_scale, --num_inference_steps, --seed: sampling settings; we use 50 steps, seed 0, and cfg 5.0 on human video / 3.0 on robot data.

📦 Checkpoints

• --ckpt: the Ropedia W0 checkpoint (a full DiT state dict, --ckpt_type full).
• --base_model_dir: local download of Wan-AI/Wan2.2-TI2V-5B (default models/Wan2.2-TI2V-5B).

A clip directory holds:

Batch inference over a CSV (--data_root, --metadata, --idx) loads the model once per process; see python inference_5b.py --help.

• Code: Apache License 2.0.
• Model weights: CC BY-NC 4.0.
• Xp-Teleop data: CC BY 4.0.

The Wan2.2 base model and the source datasets of the example clips remain subject to their original licenses.

@article{liu2026ropedia,
  title={Ropedia W0: Scaling Loco-Manipulation World Models across Embodiments with Human Experience},
  author={Liu, Tianqi and Zhao, Hanfei and Liang, Qifeng and Pang, Hui En and Xu, Jiarui and Mu, Sayang and Qiu, Penghao and Zhang, YingSheng and Wang, Jun and Cao, Ziang and Sun, Mingze and Cao, Yukang and Hong, Fangzhou and Chen, Zhaoxi and Liu, Ziwei},
  journal={arXiv preprint arXiv:2610.26855},
  year={2026}
}

We thank the authors of the following projects and datasets.