Peking University's DAGroup published Open-HumanWAM, an open recipe to pretrain world–action models on fifty thousand hours of egocentric human video and post-train them into robot policies that lead reported RoboTwin 2.0 numbers among compared open checkpoints. The arXiv paper ( 2611.06817 ), project site, Hugging Face models, and dataset card went live in early November 2026.
If you follow embodied AI because human egocentric video is the scalable manipulation prior, this release is one of the first to bundle corpus, checkpoints, training code, and benchmark scripts under a single branding umbrella rather than a leaderboard screenshot alone.
TL;DR #
| Question | Answer |
|---|---|
| Corpus | Open-HumanWAM-50K — curated egocentric video, hand motion + language |
| Models | HumanWAM-Vanilla (human-only pretrain) andHumanWAM-Goal (human + robot, visual goals) |
| Key sim result (post-trained) | Up to ~88.65% average onRoboTwin 2.0 Clean + Random per project tables |
| Scaling claim | Log-linear link between pretraining hours and downstream success; diversity matters |
| Open assets | GitHub HumanNet branch , HF weights + dataset |
| Builder angle | Reproduce scaling curves before betting your robot stack on proprietary human-video APIs |
What the paper adds beyond another VLA headline #
The abstract on Hugging Face Papers states the problem bluntly: raw human video scale does not automatically become generalizable robot policies without curation and a training recipe. Open-HumanWAM claims three contributions readers can verify if they download the assets:
- A 50k-hour corpus with dense hand-motion and natural-language annotation—not a token YouTube scrape without labels.
- Systematic scaling experiments in simulation that produce an empiricallog-linear law between human-video hours and robot success after adaptation.
- Two model lines —HumanWAM-Vanilla (human pretrain only) and HumanWAM-Goal (joint human–robot with visual goals)—withopen weights and evaluation harnesses.
HumanWAM-Vanilla, after robot post-training, is reported at 92.08% success on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random, with additional RoboDojo metrics in the paper text (about 14.17% mean success and 19.00 mean progress score in the HF summary). HumanWAM-Goal pushes the harder randomized split to 84.60%, averaging 88.65% across Clean and Random in the post-trained regime described on the project site.
Those numbers compete with other 2026 world–action releases we have covered—Verne V0.5 reported 81.0% / 73.4% clean/random on RoboTwin with an 8B checkpoint, and SLA continual pretraining reported 72.3% Clean2Random. Do not treat leaderboard ranks as eternal—each paper uses slightly different post-training budgets—but Open-HumanWAM's selling point is that you can download their checkpoints and rerun the same RoboTwin scripts.
HumanWAM-Vanilla: scaling human experience #
HumanWAM study protocol from egocentric pretrain through RoboTwin adaptation
*Image: [PKU DAGroup](https://dagroup-pku.github.io/open-humanwam/?ref=explainx)*
The [project microsite](https://dagroup-pku.github.io/open-humanwam/?ref=explainx) walks through a three-step protocol:
- Learn from people — video, language, and hand motion in egocentric frames.
- Adapt to robots — post-train on50 RoboTwin tasks with50 clean demonstrations each.
- Test generalization — clean scenes, scene out-of-distribution, and randomized clutter.
Scaling charts compare 100 → 10,000 → 50,000 hours under both isolated attention and inverse-dynamics attention variants. The authors emphasize gains outside training scenes, not only memorizing tabletop layouts.
Inverse-dynamics attention scaling from 100 to 50,000 hours of egocentric pretraining
Image: PKU DAGroup A second study splits scaled-diversity corpora (more tasks and scenes as hours grow) from coverage-matched corpora (more trajectories within fixed task/scene coverage). The site argues diversity scaling produces steeper transfer than repeating similar clips—a lesson for teams paying for egocentric annotation APIs without widening scene coverage.
Scaled-diversity versus coverage-matched corpus scaling curves
*Image: [PKU DAGroup](https://dagroup-pku.github.io/open-humanwam/?ref=explainx)*
## HumanWAM-Goal: visual goals and zero-shot behavior
HumanWAM-Goal trains a visual-goal-conditioned world–action model on hybrid human and robot data. Before RoboTwin-specific post-training, the project lists 13.94% average zero-shot success across Clean and Random, which they describe as leading among successfully evaluated open-source zero-shot checkpoints in their comparison table—not the same protocol as the post-trained 88.65% figure.
On real hardware, the site reports 20.0% → 30.5% average success across twenty tasks when moving from robot-only pretraining to hybrids with 15.9k and 45.9k hours of human data, with robot experience fixed at 14,618 hours in that ablation.
Qualitative demos show goal-conditioned in-context learning: the policy keeps weights frozen while the visual target changes—useful for teams thinking about Natura-1-style dexterous transfer without collecting hundreds of new teleop demos per SKU.
Long-horizon clips include desk organization (fourteen stages) and mobile tidying on Astribot S1, plus cross-embodiment tests where a human-view goal image guides a different robot embodiment after task post-training. Treat those videos as authors' demonstrations until independent labs reproduce them.
How to try it this week #
Read and cite
- Paper PDF: [arxiv.org/abs/2611.06817](https://arxiv.org/abs/2611.06817?ref=explainx)
- Overview and films: [dagroup-pku.github.io/open-humanwam/](https://dagroup-pku.github.io/open-humanwam/?ref=explainx)
Download
- Models collection: `huggingface.co/collections/DAGroup-PKU/open-humanwam`
- Dataset: `huggingface.co/datasets/DAGroup-PKU/Open-HumanWAM-50K`
- Code: `github.com/DAGroup-PKU/HumanNet` branch`open-humanscale`
Evaluate
- Start in RoboTwin 2.0 with their published post-training recipe before attempting real robots.
- Log Clean vs Random separately; Random is whereXiaomi U0-style synthesis and domain-randomization papers claim the real generalization win.
Compute expectations
Fifty thousand hours of video is not a laptop fine-tune. Plan cluster storage for the corpus subset you need, mirror HF artifacts regionally, and budget GPU time for world-model pretraining at the scales their curves imply—100-hour baselines are useful sanity checks, not the headline model.
RoboTwin 2.0 and RoboDojo in plain language #
RoboTwin 2.0 is a two-arm manipulation benchmark with fifty tasks, each trained from tens of clean demonstrations, then tested in matched clean scenes and randomized scenes where object poses, textures, and lighting shift. Papers report Clean2Clean and Clean2Random separately because randomized success collapses when policies memorize a single tabletop layout.
Open-HumanWAM's ~80% vanilla / ~89% goal averages are post-trained numbers: they assume you run the authors' adaptation stage on RoboTwin demonstrations after human-video pretraining. The 13.94% zero-shot figure for HumanWAM-Goal is a different protocol—evaluating the pretrained checkpoint before task-specific post-training—so it should not be compared directly to the eighty-eight percent line in the same breath.
RoboDojo metrics in the paper (mid-teens percent success in the HF abstract) act as a secondary stress test. Treat them as additional evidence, not the headline marketers will tweet.
World–action models versus classic VLAs #
Classic vision-language-action (VLA) stacks map an image and instruction directly to joint targets or end-effector deltas. World–action models (WAMs) also model short-horizon dynamics or video prediction jointly with actions, which can help when contact-rich physics matter.
HumanWAM's architecture diagrams on the project site emphasize closed-loop simulation evaluation during scaling studies—that is, the model is not judged only on next-token loss on human YouTube clips. For builders comparing stacks, our LLM vs CLM vs decision models guide is about language heads, not robot WAMs, but the same lesson applies: match the evaluation harness to the deployment harness before choosing a pretrained initialization.
If you already run a π₀-class diffusion policy or a smaller 8B Verne checkpoint, treat HumanWAM as a pretraining initializer experiment: swap the human-video encoder weights, keep your robot data collection pipeline fixed for a month, and measure data efficiency (demos needed to hit 70% on your task) rather than chasing leaderboard points on tasks you will never ship.
What people are asking on Hugging Face #
The author comment on the paper page summarizes HumanWAM-Vanilla at 92.08% / 68.88% Clean/Random (~80.48% average) and HumanWAM-Goal at 92.70% / 84.60% (~88.65% average) after post-training, plus RoboTwin rank #1 language in their self-reported comparison.
A community comment from PolyHelperAI notes the log-linear hours-to-success relationship and asks where the curve flattens—a fair open question the paper's appendices may address with larger held-out suites. Treat flattening points as empirical, not guaranteed log-linear forever.
The automated Librarian Bot thread links adjacent 2026 papers (IronMind, InternW0, Ego-CEVR, and others). Readers building literature reviews should cluster those works on egocentric pretraining rather than reading Open-HumanWAM in isolation.
Limits and honest comparisons #
Simulation-first. Most cited percentages are RoboTwin 2.0 or simulation analogs. Real-robot zero-shot averages in the teens-to-thirties percent range on the project site are far from production reliability.
Naming drift. Checkpoints use the HumanWAM prefix; the Git branch is open-humanscale while Hugging Face collections say open-humanwam. The corpus card is Open-HumanWAM-50K everywhere official—do not confuse it with the separate HumanNet megacorpus roadmap.
Leaderboard volatility. Verne's model cards already showed rivals ahead on some RoboTwin sorts while marketing claimed leadership. Always read which split, which post-training budget, and which open-vs-closed row a table includes.
Data ethics and consent. Fifty thousand hours of egocentric human video raises privacy and consent questions the technical paper may only partially document. Teams fine-tuning in the EU or California should trace dataset provenance fields on Hugging Face before merging clips into commercial robot products.
Where this fits the embodied foundation model map #
Open-HumanWAM sits in the open recipe camp with GreenWAM-style releases (Sber GreenWAM) and closed-stack giants shipping APIs only. The differentiator is reproducible scaling science: curves tying hours to success, ablations on diversity, and zero-shot goal conditioning you can inspect—not only a mean success scalar.
For practitioners, the actionable bet is empirical: if your robot line already pays for human video, compare fine-tuning with and without a HumanWAM initialization before commissioning another year of teleop. If you are simulation-only, rerun their 100-hour vs 1k-hour checkpoints to see whether log-linear gains appear on your end-effector and camera mount. Industrial labs chasing reward-guided OM-1 style human data should watch whether DAGroup publishes training cost envelopes in later revisions; without them, total cost of ownership vs proprietary APIs stays fuzzy.
Related reading #
Benchmark numbers, dataset availability, and repository branches reflect DAGroup-PKU releases as of November 9, 2026.