NTH

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

AuthorsJuncheng 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 6, 2026 2 min read
Watch on YouTube
The one-line take

Open-HumanWAM scales robot manipulation learning with 50,000 hours of human video and shared human–robot training.

Key results

50K
Open-HumanWAM-50K duration

Hours of curated egocentric human video in the corpus.

35.9M
Open-HumanWAM-50K clips

Continuous clips retained after curation.

92.70%
HumanWAM-Goal Clean2Clean

RoboTwin 2.0 success after robot post-training.

84.60%
HumanWAM-Goal Clean2Random

RoboTwin 2.0 success after robot post-training under randomized conditions.

30.5%
Physical-robot zero-shot success

Mean success across 20 tasks with 14.6K robot hours and 45.9K added human hours.

45.9K
Added human pretraining data

Hours added to the fixed robot corpus in the physical-robot zero-shot scaling experiment.

What the paper found

Open-HumanWAM presents an open recipe for teaching robots from human egocentric video, with Open-HumanWAM-50K: a curated corpus of 50K hours and 35.9M clips annotated with hand motion and language, aligned to a shared action representation. The study finds that scaling human data improves downstream out-of-distribution robot success, and that expanding task and scene diversity strengthens the trend. Its HumanWAM-Vanilla model is pretrained only on human video, while HumanWAM-Goal combines human and robot experience, adds visual-goal conditioning, and uses a 49-dimensional hybrid action interface. Built on NVIDIA’s Cosmos3-Nano-16B, HumanWAM-Goal reaches 92.70% on RoboTwin 2.0 Clean2Clean and 84.60% on Clean2Random after robot post-training. On physical robots, adding 45.9K hours of human video to a fixed 14.6K-hour robot corpus raises zero-shot success across 20 tasks to 30.5%; the frozen policy can also change its behavior in response to different goal images, without parameter updates. OpenAI’s GPT Image 2.5 and ChatGPT Astra High are used to generate and review simulated goal images.

Original 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.

Read the original paper

More in Embodied AI

Browse all 69 papers →
02Embodied Ai

Kimodo-Policy: Learning Kinematic Action Models for Humanoid Loco-Manipulation

Yunheng Wang, Jiaxi Zhang, Wei Lu, Yuetong Fang, Taowen Wang, Yixiao Feng, Shuning Zhang, Junzhe Xu, Yichi Wang, Zizhao Yuan, Kailun Huang, Lusong Li, Mingqian Zhou, Zecui Zeng, Renjing Xu

Kimodo-Policy turns a pretrained whole-body motion generator into a control policy for humanoid robots that walk and manipulate objects.

Read analysis