Open-HumanWAM: An Open Recipe for Scaling Human and Robot Experience for Embodied Foundation Models

Juncheng Ma, Jiaxin Li, Jiankai Tu, Bo Liang, Guodong Lv, Yufan Deng, Junzhe Tang, Liangzhu Yang, Chen Bian, Shukai Gong, Yintianrun Zhang, Zicong Cao, Baohao Zhu, Jianxin Bi, Tengfei Yan, Yinyuan Zhang, Kewei Zhang, Xinglong Chen, Kaiqi Chen, Duomin Wang, Xiaotian Tang, Liang Tan, Li Li, Yuan Xu, Haosen Yang, Yuqi Wang, Jingfei Xia, Songyan Huang, Zhiyang Dou, Zhen Dong, Enze Xie, Tat-Seng Chua, Daquan Zhou

November 2026

The Astribot S1 filing notebooks and handling a mug on an office desk.

TL;DR A world-action model pretrained on a curated 50,000-hour corpus of egocentric human video, with robot success after post-training rising log-linearly in pretraining hours. On a real bimanual robot, mean success under new tablecloths and lighting rises from 2.5% without human pretraining to 67.5%.

What they do

An open recipe for pretraining robot policies on egocentric human video. The authors curate Open-HumanWAM-50K (50,000 hours from five public datasets, with hand-motion and language labels) and measure robot success as pretraining grows from 100 to 50,000 hours. Two models result: HumanWAM-Vanilla, pretrained on human video only, and HumanWAM-Goal, co-trained on human and robot data and conditioned on a goal image.

How it works

Curation and alignment. Clips are filtered for hand visibility, smooth pose tracks and actual manipulation. Wrist poses from every source go into one 16-D bimanual end-effector space: camera frame, canonical wrist axes, and gripper openness taken from the thumb-index distance.
HumanWAM-Vanilla. Wan2.2-5B video model plus an action expert. It predicts future video and actions; with inverse-dynamics attention, actions attend to the predicted future frames.
HumanWAM-Goal. Built on Cosmos3-Nano-16B with a 49-D action space covering joints, base and end-effectors. Trained on 14.6k robot hours plus up to 45.9k human hours, with goal images sampled from future frames.

Results

RoboTwin 2.0 success follows SR = a + b log10(hours). From 100 to 50k hours, scene-ood success rises from 49.1% to 73.2% and random from 21.5% to 42.9%. With 1k hours, human pretraining beats 5k hours of robot pretraining.
Real Cobot Magic, 4 tasks: in-distribution success 57.5% → 87.5%. Under new tablecloth and lighting it goes 2.5% → 67.5%.
Real Astribot S1, 20 tasks, zero-shot: 20.0% robot-only, 30.5% with 45.9k human hours. A 1:1 robot-human batch ratio raises hard-task success from 19% to 27%.
After post-training, HumanWAM-Goal reaches 92.70% Clean / 84.60% Random on RoboTwin 2.0, with extra augmentation beyond the scaling recipe.

Limitations

The paper has no limitations section. HumanWAM-Vanilla trails the strongest RoboDojo entries (14.17% mean success); the authors attribute this to having no memory or reasoning modules. Zero-shot success on hard real tasks stays below 30%. In-context goal following and long-horizon runs are shown only as qualitative rollouts.

Videos

Head-camera view: a person moves the bottle, then the robot works on the same scene.

Head-camera view: a person moves a blue toy, then the robot picks it up.

Figures

From Figure 1: Overview of Open-HumanWAM. Scaling a curated 50,000-hour egocentric corpus improves robot generalizati... — From Figure 1: Overview of Open-HumanWAM. Scaling a curated 50,000-hour egocentric corpus improves robot generalization after post-training, with more task and scene diversity giving steeper scaling curves. Combining human and robot data improves zero-shot success in simulation and across 20 real-world tasks.

From Figure 3: Model architecture of HumanWAM-Vanilla and experimental protocol. (a) Egocentric pretraining, robot po... — From Figure 3: Model architecture of HumanWAM-Vanilla and experimental protocol. (a) Egocentric pretraining, robot post-training on clean demonstrations, and closed-loop evaluation under clean, scene-ood and random conditions. (b) The video and action experts interact through joint masked self-attention; the masks set which current-frame, future-video and action tokens see each other under isolated and inverse-dynamics attention.

From Figure 5: Scaling egocentric pretraining with isolated attention. Blue dashed lines are least-squares fits of su... — From Figure 5: Scaling egocentric pretraining with isolated attention. Blue dashed lines are least-squares fits of success rate = a + b log10(H), where H is pretraining hours, with R2 in each panel; gray dashed lines are the baseline without egocentric pretraining. Success rises in all three settings, with larger gains on scene-ood and random.

Abstract

Egocentric human video offers a scalable source of manipulation priors for robot learning, yet translating raw data scale into generalizable robot policies demands rigorous data curation and training recipes. We introduce Open-HumanWAM, the first fully open recipe for scaling egocentric pretraining to 50,000 hours, anchored by Open-HumanWAM-50K---a curated egocentric video corpus densely annotated with hand motion and natural language. Leveraging this corpus, we systematically investigate pure human pretraining via closed-loop simulation evaluations and uncover an empirical log-linear scaling law linking pretraining data hours to downstream robot success rates. Grounded in these findings, we develop HumanWAM-Vanilla, a world--action model pretrained exclusively on 50k hours of egocentric human video. Upon robot post-training, HumanWAM-Vanilla reaches 92.08% success on RoboTwin 2.0 Clean2Clean and 68.88% on Clean2Random, alongside a 14.17% mean success rate and a 19.00 mean progress score on RoboDojo. To push the capability frontier further, we explore joint human--robot pretraining with HumanWAM-Goal, a visual-goal-conditioned world--action model. Hybrid data scaling unlocks broad zero-shot generalization across both simulation and physical benchmarks, propelling post-trained RoboTwin 2.0 performance to new state-of-the-art levels: 92.70% on Clean2Clean and a substantial leap to 84.60% on Clean2Random. On physical robots, HumanWAM-Goal further exhibits emergent embodied in-context learning, dynamically adapting its manipulation behavior to unseen visual goals without parameter updates. To support reproducible scaling research in embodied AI, we open-source the Open-HumanWAM-50K corpus, data curation and pretraining pipelines, alongside the pretrained and post-trained HumanWAM checkpoints, inference code, and simulation evaluation suites.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong