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

1. AnnotationLabel with humans in the loop

   Automated models annotate each clip; targeted human checks ensure quality.

2. Public dataRe-annotate public data

   14 human and 13 robot datasets, unified in one schema.

3. Our recordingsRecord our own

   6,000+ hours of multi-view egocentric video from a four‑fisheye rig.

4. BenchmarkBuild Xp-Teleop

   Whole-body dexterous humanoid teleoperation in six household activity groups, for long-horizon evaluation.

1. Locomotion
   Camera poses

   Relative to the first frame, as dense Plücker-ray maps; every step and turn moves the view.

2. Manipulation
   Projected skeletons

   Hand keypoints projected into each frame, from pose annotations for humans and forward kinematics for robots.

3. Scale
   Two model scales

   A dense 5B model and A14B, a mixture-of-experts with 14B active parameters.

4. Training
   From humans to robots

   Human pre-training, then human–robot mid-training; optional post-training on a target robot.

1. HumanBetter human prediction

   Camera and hand following improve steadily as human pre-training grows.

2. RobotBetter robot prediction

   The gains carry over to robot prediction across embodiments.

1. BenchmarksLead across benchmarks

   State of the art on human video and every robot benchmark, from zero-shot to full-data post-training.

2. BaselinesOutperform Cosmos 3

   Matches the ground truth more closely than both Cosmos 3 models, including the larger Cosmos 3-Super.

3. EmbodimentsCover many embodiments

   Humans, gripper arms and dexterous humanoids, steady as the body walks and turns.

4. Long horizonLead on long rollouts

   Stays ahead of the baselines over long rollouts that chain locomotion and manipulation.

Lead across benchmarks

Quantitatively, Ropedia W0 leads on held-out human video and five robot benchmarks (DreamDojo, Ψ0, ω-HOME, AgiBot and our Xp-Teleop), from zero-shot to full-data post-training; on DreamDojo's own protocol, it leads at every data budget.

DreamDojo benchmarkAll benchmarks

DreamDojo benchmark · GR-1 post-training

PSNR (dB) ↑

1%10%100%

Swipe to see all benchmarks →

• DreamDojo (2B)
• DreamDojo (14B)
• Ropedia W0 (5B)
• Ropedia W0 (A14B)

All benchmarks · zero-shot to full data

Human video and four robot benchmarks; further out is better

• Ropedia W0 (A14B)
• Ropedia W0 (5B)
• Cosmos 3-Nano (16B)
• Cosmos 3-Super (64B)
• DreamDojo (14B)

Ropedia W0 (A14B) is the outermost line on every axis.

Compare with state-of-the-art world models

Qualitatively, given the same input frame and actions, Ropedia W0 stays closer to the ground truth than Cosmos 3.

One model, many embodiments

A shared kinematic action space lets one model predict humans, gripper robots and dexterous humanoids, with consistent scenes under locomotion.

Predict over long horizons

Long autoregressive rollouts chain locomotion and manipulation, each generated from a single real frame and the commanded actions.

1. r ≥ 0.98Policy evaluation

   Correlation between imagined and real task progress; the checkpoint ranking is preserved.

2. +7 / +10 ppPlanning

   Over the best single checkpoint, in simulation / on a real robot, by planning in imagination.

3. 21.3 fpsReal-time teleoperation

   640×480 video streamed on one GPU by a 4-step distilled model, for live VR control.

4. 33.3 → 56.7%World–action models

   Real Unitree G1 task progress: FastWAM vs. FastWAM initialised from Ropedia W0.

Evaluate and plan in imagination

As a proxy environment, Ropedia W0 ranks policy checkpoints in the same order as real runs (MMRV = 0). For planning, it imagines candidate action sequences, a value model scores them, and the best one is executed.

Teleoperate live in VR

Distilled into a 4-step autoregressive model, Ropedia W0 streams 640×480 video at 21.3 fps on one GPU and stays stable for minutes. An operator in a Quest 3 headset moves, and the model renders what a human or robot body would see.

Learn better policies

Initialising the video backbone of a world–action model (FastWAM) from Ropedia W0 5B, with everything else unchanged, raises held-out LIBERO-Plus success from 50.1% to 56.8% and helps every few-shot setting. On a real Unitree G1 the resulting policy reaches 56.7% task progress, against 33.3% for FastWAM; its executions are shown below.

@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={Technical report},
  year={2026}
}