HeyHomees Lab Notes · No. 6 · Ropedia · Published October 21, 2026
By HeyHomees Newsroom ·
In 30 seconds
- A single 5 billion-parameter model learns loco-manipulation from 10 k human and 3 k robot hours of video.
- It achieves state-of-the-art results on five cross-embodiment benchmarks and held-out human video.
- Runs in real-time on an NVIDIA GPU with ≥40 GB memory, using 640×480 resolution and 81-frame rollouts.
What is it?
Ropedia W0 is an egocentric world foundation model that predicts the consequences of locomotion and manipulation actions across multiple embodiments, including humans, robot arms, humanoids with grippers, and dexterous hands. It is trained on a unified dataset of egocentric video with camera poses and projected kinematic skeletons. The model is released in a 5 billion-parameter version.
How does it work?
The system encodes locomotion via camera poses and manipulation via projected skeletons, using robot-specific forward kinematics combined with manually annotated keypoints. Training data comprise 10 k hours of human video and 3 k hours of robot video, all with metric camera trajectories, skeletons, and hierarchical captions. After pre-training, the model can be distilled for real-time, minute-long rollouts.
How well does it do?
Evaluations report state-of-the-art performance on five cross-embodiment benchmarks (DreamDojo, Ψ₀, ω-HOME, Xp-Teleop, and AgiBot) and on held-out human egocentric video, covering zero-shot, few-shot, and full-data post-training settings. Scaling the amount of human pre-training data consistently improves predictions for both humans and robots.
What does it need to run?
Running the 5B model requires Linux, an NVIDIA GPU with at least 40 GB of memory (31 GB peak usage at 640×480 resolution with 81 frames), and CUDA 12.8 (or CUDA 13 drivers). The code uses Python 3.12, Conda, and PyTorch 2.11 with SDPA; no custom CUDA extensions are needed.
By the numbers
- Human video data: 10k hours
- Robot video data: 3k hours
- Total unified experience: 13K Hours
- GPU memory requirement: at least 40 GB
- Peak memory usage: 31 GB
- Input resolution and frames: 640×480, 81 frames
The honest read
The reported state-of-the-art results are based on the authors' own evaluations on five robot benchmarks and held-out human video, so the scores are self-reported and have not been independently verified. The performance gains are described as consistent when increasing human pre-training data, but no absolute numbers or external baselines are provided, limiting assessment of the magnitude of improvement. Real-time operation is claimed after model distillation, yet the required GPU memory (≥40 GB) may restrict deployment to high-end hardware.
What we don't know yet
- Exact benchmark scores or comparisons to prior models are not provided.
- The download size of the model checkpoint is not stated.
- Licensing terms for the model and data are not specified.
Sources
- Paper (arXiv abstract) (primary)
- Code, GitHub
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.