HumanWAM

HumanWAM (Human World–Action Model) is the embodied world–action model family from DAGroup-PKU and collaborators in the Open-HumanScale project: pretrain on large-scale egocentric human manipulation video, then adapt to bimanual robots via post-training. Paper: https://arxiv.org/abs/2611.06817 (arXiv 2611.06817); HF example: https://huggingface.co/DAGroup-PKU/HumanWAM-Vanilla-Base ; project: https://dagroup-pku.github.io/open-humanscale/ .

Data & lines:

• Open-HumanScale-50K: ~50,000 hours of curated egocentric manipulation video with hand-motion and language annotations;
• HumanWAM-Vanilla: pretrained exclusively on human data; the HF Vanilla-Base release includes a human egocentric pretraining checkpoint and an EgoPro multi-view mid-training checkpoint for robot post-training init;
• HumanWAM-Goal: joint human–robot pretraining with visual-goal conditioning for zero-shot execution and embodied in-context learning.

Paper-reported post-training results: Vanilla reaches 92.08% Clean2Clean and 68.88% Clean2Random on RoboTwin 2.0; Goal reaches 92.70% / 84.60% (average 88.65%). On four real-robot tasks under visual OOD, mean success rises from 2.5% without egocentric pretraining to 67.5% with Vanilla.

Problem

Robot foundation models need diverse manipulation experience, but teleop data is costly. Egocentric human video scales manipulation priors; Open-HumanScale open-sources the corpus, curation/alignment recipe, and an empirical log-linear scaling law between pretraining hours and downstream success.

Model lines

• Vanilla: human-only world–action pretraining (joint future video + actions), then mid-train / post-train for multi-view robot policies;
• Goal: hybrid human–robot data plus goal-image conditioning for zero-shot control and parameter-free visual in-context adaptation.

Release status & limits

• HF Vanilla-Base is weights-first: pretrained/mid-trained weights are up; matching normalization stats, configs, and runtime resources land separately—downloads alone are not a full runnable package;
• Task robot inference needs the matching post-trained policy and runtime assets;
• Benchmark numbers are from the paper; reproduce with the official eval suites.

Who it's for

• Embodied AI / robot-learning researchers needing an open human-pretraining recipe and checkpoints;
• Teams post-training for RoboTwin or real bimanual setups.

Getting started

1. Read https://arxiv.org/abs/2611.06817 and the project page (Vanilla vs Goal);
2. hf download DAGroup-PKU/HumanWAM-Vanilla-Base --local-dir HumanWAM-Vanilla-Base;
3. Use https://github.com/DAGroup-PKU/HumanNet/tree/open-humanscale and the HF Open-HumanScale collection for code/data;
4. Obtain normalization/runtime assets before full closed-loop inference—do not treat Base weights alone as a finished policy.