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.
Figures & tables
Figure 1: Overview of Open-HumanWAM. Using a curated 50,000-hour egocentric corpus, we show that scaling human data improves robot generalization after post-training, with greater task and scene diversity yielding steeper scaling curves. We then extend this pretraining approach to combine human and robot data, finding that hybrid scaling improves zero-shot success in simulation and across 20 real-world tasks.
Figure 2: Construction of Open-HumanWAM-50K. (a) Five data sources and their retained durations. (b) Data curation based on hand visibility, motion stability, and manipulation activity. (c) Cross-source alignment of camera frames, wrist axes, and gripper openness.
Dataset
Before curation
After curation
Hours
Frames (M)
Hours
Frames (M)
Episodes
HumanNet subset [ 11 ]
90,275.25
9,749.73
40,743.58
4,400.31
30,833,247
EgoSuite-Open100K [ 75 ]
11,193.99
1,208.95
7,081.65
764.82
348,724
EgoDex [ 20 ]
829.45
89.58
335.43
36.23
163,537
Gen-HumanEgo [ 17 ]
1,847.39
199.52
1,814.53
195.97
44,632
EgoVerse [ 43 ]
143.33
15.48
24.81
2.68
6,659
Table 1: Source datasets and curation statistics for Open-HumanWAM-50K
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) Conditioned on language and proprioception, the video and action experts interact through joint masked self-attention. The masks specify visibility among current-frame (C), future-video (F), and action (A) tokens under isolated attention (IA) and inverse-dynamics (IDM) attention.
Figure 4: RoboTwin evaluation setup for our egocentric pretraining exploration. We post-train policies exclusively on clean demonstrations and evaluate them under clean, scene-ood, and random conditions.
Figure 5: Scaling egocentric pretraining with isolated attention improves in-distribution performance and generalization. Blue dashed lines show least-squares fits of SR=a+blog10H , where H denotes pretraining hours, with R2 reported in each panel. Gray dashed lines denote the baseline without egocentric pretraining. Success rates increase across all three settings, with larger gains on scene-ood and random, indicating that scaling primarily strengthens generalization beyond the demonstration scenes.
Figure 6: Scaling egocentric pretraining with IDM attention strengthens downstream generalization. Clean success remains high across scales, while scene-ood and random success improve consistently. Blue dashed lines show log-linear scaling fits for the two out-of-distribution settings, with R2 reported in the corresponding panels. Gray dashed lines denote the baseline without egocentric pretraining.
Family
Hours
Clips
Tasks
Scenes
Scaled-diversity
100
62,206
62
222
500
305,889
309
961
1,000
589,216
547
1,442
Matched-diversity
100
81,076
2,288
2,953
500
345,053
2,288
2,953
1,000
653,567
2,288
2,953
Table 3: Task and scene coverage of the two EgoSuite-Open100K corpus families and their shared 5,000-hour parent corpus. Task and scene counts are based on the dataset captions.
Figure 7: Scaling egocentric pretraining with growing or fixed task and scene diversity. Blue circles denote scaled-diversity corpora, and orange squares denote matched-diversity corpora. Dashed and dash-dotted lines show the respective least-squares fits of SR=a+blog10H , with both equations shown in each panel. Scaled-diversity yields a steeper fitted slope in all three settings.
Figure 8: Scaling egocentric pretraining improves the sample efficiency of robot adaptation. Success rates as a function of the number of robot demonstrations per task for different egocentric pretraining scales.
Figure 9: Cross-source action alignment improves downstream policy performance. Aligned and Raw models are pretrained on the same four-source 1,000-hour corpus using aligned and native action representations. Each point reports the best success rate across post-trained checkpoints for the corresponding pretraining run and evaluation setting. Gray dashed lines denote the baseline without egocentric pretraining.
Figure 10: Cross-source action alignment reduces pretraining action loss with little change in video loss. Curves show training losses for Aligned and Raw models on the same four-source 1,000-hour corpus.
Figure 11: Human action supervision during egocentric pretraining improves downstream generalization. Models are pretrained on the same 1,000-hour video corpus with 0%, 50%, or 100% human action supervision.
Scale (h)
AgiBot Beta (h)
XDOF/ABC-130k (h)
Cobot (h)
Total frames
100
32.19
67.80
—
10,798,713
500
176.52
323.46
—
53,997,982
1,000
361.70
638.16
—
107,985,621
5,000
1,944.80
2,983.46
71.52
539,977,159
Table 4: Composition of the real-robot pretraining corpora.
Figure 12: Egocentric pretraining outperforms robot pretraining from 500 to 5,000 hours. Downstream RoboTwin success rates after pretraining on 100, 500, 1,000, and 5,000 hours of egocentric human or robot data. The egocentric series follows the scaled-diversity configuration in Figure 5 . Each point reports the best post-training success rate among the evaluated pretraining checkpoints for that scale and setting.
Figure 13: From egocentric pretraining to multi-view robot policies. Three view configurations and training routes with or without EgoPro mid-training.
Model configuration
View configuration
Mid-training
Clean
Scene-OOD
Random
Isolated attention (relative)
Single-view
None
75.68
46.08
22.48
Multi-view conditioning
None
74.60
34.08
12.92
Multi-view prediction
None
75.04
33.00
14.84
IDM attention (absolute)
Single-view
None
88.48
62.00
28.56
Multi-view conditioning
None
90.56
59.52
29.08
EgoPro
90.28
61.20
30.68
Table 5: From egocentric pretraining to multi-view robot policies. RoboTwin success rates (%) from the best available complete post-training evaluation for each setting.
Setting
Robot demonstrations
Optimizer steps
Global batch size
RoboTwin 2.0
2500
24,000
512
RoboDojo
3500
50,000
512
Real robot
50/task
10,000
256
Table 6: Post-training settings for RoboTwin 2.0, RoboDojo, and real-robot experiments.
Type
Method
Clean2Clean ↑
Clean2Random ↑
Average ↑
VLA
Qwen-RobotManip-Context [ 66 ]
84.7
69.4
77.1
GigaBrain-0.7 [ 51 ]
66.80
67.92
67.36
π0.5 [ 21 ]
70.70
46.00
58.35
Spatial Forcing [ 24 ]
77.20
26.74
51.97
X-VLA [ 72 ]
68.00
20.90
44.45
Abot-M0 [ 62 ]
57.40
30.40
43.90
Table 7: Evaluation results on RoboTwin 2.0. Success rates (%) are reported after clean-only post-training. Average is the mean of Clean2Clean and Clean2Random. The best result in each column is highlighted in bold.
Figure 14: Real-robot tasks and visual evaluation conditions.
In-distribution
Out-of-distribution
Task
No-Ego
HumanWAM-Vanilla
No-Ego
HumanWAM-Vanilla
Paper cup stacking
70.0
90.0
0.0
80.0
Towel folding
50.0
80.0
10.0
60.0
Drawer storage
60.0
90.0
0.0
70.0
Desk tidying
50.0
90.0
0.0
60.0
Mean
57.5
87.5
2.5
67.5
Table 9: Real-robot task success with and without egocentric pretraining. Success rates are reported in percent over 10 trials per task, model, and evaluation setting. Mean averages the four task success rates.
Attention
Action representation
Inference
Clean
Scene-OOD
Random
IA
Relative
Action-only
76.28
46.36
22.40
Mutual
Relative
Joint denoising
81.28
57.96
32.92
IDM
Relative
Video → action
83.56
58.08
27.56
Mutual
Absolute camera-frame
Joint denoising
84.88
61.12
30.68
IDM
Absolute camera-frame
Video → action
88.48
62.00
28.56
Optional-IDM
Relative
First-frame only
73.44
48.44
23.40
Table 10: Ablation of attention strategies and action representations.
Action-expert initialization
Clean
Scene-OOD
Random
Interpolated Wan2.2 weights
89.56
83.20
51.52
Random initialization
90.28
83.92
54.12
Table 11: Ablation of action-expert initialization.
Figure 15: Robot pretraining data overview. The top row reports the overall robot pretraining budget. Word clouds summarize selected (a) action and (b) object terms, while the adjacent bars show label-level percentages for the eight most frequent terms in each group. Percentages use all 751 normalized task-collection labels as the denominator.
Figure 16: HumanWAM-Goal architecture. (a) Conditioning. An explicit goal image augments language instructions, current and optional historical observations, and the 49-D robot state. (b) Joint rectified-flow sampling. With conditioning tokens held fixed, the shared video-action Generator iteratively predicts video and action velocities via UniPC updates.
Setting
Target-task adaptation
Episodes
Results
Zero-shot
None
50
Table 13
Post-training
50 clean demos/task
100
Figure 19
Leaderboard
Best configuration + augmentation
100
Table 14
Table 12: RoboTwin 2.0 evaluation protocols.
Model
Clean
Random
Average
LingBot-VLA 2.0
ERR0
ERR0
ERR0
LingBot-VA
ERR0
ERR0
ERR0
OpenWAM [ 56 ]
ERR0
ERR0
ERR0
DM0.5
ERR1
ERR1
ERR1
GR00T N1.7
ERR1
ERR1
ERR1
DreamZero-AgiBot [ 65 ]
ERR1
ERR1
ERR1
Table 13: Zero-shot evaluation on RoboTwin 2.0 (ALOHA-AgileX). Pretrained checkpoints are evaluated directly without RoboTwin-specific adaptation using 50 episodes per task under both Clean and Random conditions. Success rates are reported in percent. Status codes ERR0 – ERR2 denote deployment barriers that prevent zero-shot evaluation.
Figure 17: Zero-shot performance under hybrid data scaling. Task success rates and task coverage ( ≥2% mean success) across six HumanWAM-Goal pretraining configurations on RoboTwin 2.0 (50 Clean and 50 Random episodes per task).
Figure 18: Task-level zero-shot scaling with human-robot hybrid training. Mean zero-shot success rates (averaged over Clean and Random) across six representative RoboTwin 2.0 tasks with robot data fixed at 14,618 hours and increasing egocentric human data.
Figure 19: Post-trained performance across pretraining scales on RoboTwin 2.0.
Type
Method
Clean2Clean ↑
Clean2Random ↑
Average ↑
VLA
Qwen-RobotManip-Context [ 66 ]
84.7
69.4
77.1
GigaBrain-0.7 [ 51 ]
66.80
67.92
67.36
π0.5 [ 21 ]
70.70
46.00
58.35
Spatial Forcing [ 24 ]
77.20
26.74
51.97
X-VLA [ 72 ]
68.00
20.90
44.45
Abot-M0 [ 62 ]
57.40
30.40
43.90
Table 14: Evaluation results on RoboTwin 2.0. Success rates (%) are reported after clean-only post-training. Average is the mean of Clean2Clean and Clean2Random. The best result in each column is highlighted in bold.
Figure 20: Zero-shot real-robot performance stratified by task difficulty. Success rates across easy and hard tasks as egocentric human data are added to a fixed 14,618h robot corpus. The vertical dashed line at 60,556h highlights the additional gain on hard tasks from robot-balanced batch sampling ( +8pp ).
Figure 21: Per-task zero-shot evaluation on physical robots. Task-level success rates (left) and mean progress scores (right) across all 20 benchmark tasks for the robot-only baseline and three human–robot hybrid configurations.
Figure 22: Qualitative real-robot zero-shot rollouts. Each row pairs the inference-time goal image (left) with four head-camera keyframes from the resulting zero-shot execution (right), showing the initial scene, two intermediate interaction stages, and the final state. The examples span object transport and placement, coordinated bimanual manipulation, object reordering, and multi-step tool/container interactions.
Figure 23: Visual goal conditioning in pretrained HumanWAM-Goal. Each panel pairs the inference-time goal image (left) with synchronized first-person and third-person keyframes (right). Without target-task post-training or parameter updates, changing the visual goal directs the policy to place the bottle (a) on the tabletop in front of the green block or (b) atop the large purple block.
Figure 24: Goal-based in-context learning after post-training for tabletop sorting and stacking. Each panel pairs a supplied goal image (left, blue border) with eight chronological third-person frames (right, read left-to-right across the top then bottom row; green border marks the final state). The goals specify (a) sorting toy cars into the blue bin, plush toys into the white bin, and stacking books near the table center with the green book on top; (b) swapping the toy-to-bin assignments; and (c) retaining the original toy assignments while placing the purple book on top of a stack shifted toward the white bin.
Figure 25: Goal-based in-context learning after post-training for office desk organization. Each panel pairs a goal image and enlarged detail view (left) with six ordered third-person execution frames (right). Panels (a, b) redirect the storage of a tape roll and tape measure from the upper drawer to the middle drawer. Panels (c, d) swap the destinations of a glue stick and two pens between the left and right holders (defined in the goal-image frame). Endpoints show item placement prior to drawer closure in (a, b) and after the final pen release in (c, d).
Figure 26: Cross-embodiment in-context learning from egocentric human goal images. Each panel pairs a human-demonstrated goal image and enlarged detail view (left, showing human hands) with synchronized first-person and third-person robot execution frames (right). Panels (a, b) direct the robot to file the brown notebook into the rightmost or leftmost rack slot, respectively. Panels (c, d) specify placing the covered mug near the monitor or beside the file rack. Green borders highlight the completed placements upon gripper release.
Figure 27: Long-horizon desk organization with HumanWAM-Goal. Fourteen chronologically ordered stages from a single continuous rollout, shown through paired first-person (left) and third-person (right) views. The trajectory sequentially executes mug capping and placement, pen and glue-stick sorting, articulated drawer storage of a tape roll and tape measure, and notebook filing. Green borders mark the terminal state after the final notebook is released into the rack.
Figure 28: Long-horizon mobile household tidying with HumanWAM-Goal. Twenty-eight chronologically ordered keyframes from a single continuous rollout spanning multiple work areas: sorting tabletop toys and books, navigating to a side table to collect and discard empty bottles, gathering garments from two chairs into a laundry hamper, and returning to the main table.
Verdict
Images
Share
Pass
6,380
90.7%
Fail
542
7.7%
Uncertain
110
1.6%
Total
7,032
100.0%
Table 15: First-pass review outcomes for the 7,032 newly generated evaluation goal images.
ID
Task
Family
Scored milestones (maximum points)
Task success outcome
Easy tasks (10 tasks; 100 trials)
E1
Lift a block
B
Approach (3) → grasp (4) → lift (3)
Target block held aloft.
E2
Compress a sponge
A
Approach (2) → contact (3) → press (5)
Sponge fully compressed.
E3
Open a hardcover book
C
Approach (2) → touch (2) → grasp (3) → open (3)
Front cover opened.
E4
Ring a desk bell
A
Approach (3) → contact (3) → press (4)
Bell actuated to ring.
E5
Press a click pen
A
Approach (3) → contact (3) → press (4)
Pen’s push button depressed.
Table 16: Real-world zero-shot tasks and scoring rubrics. Milestones are listed in the order recorded in the evaluation sheets; arrows do not impose a universal prerequisite rule. Parentheses give maximum points, not cumulative scores. E1–E10 are easy tasks; H1–H10 are hard tasks.
Embodied foundation models are expected to benefit from data scaling like large language models, but face a much tighter data bottleneck. Teleoperated real-robot trajectories remain the dominant pretraining source due to their precise action supervision and embodiment alignment, yet their scalability is limited by high collection cost, acquisition difficulty, and low behavioral and environmental diversity. These limitations have sparked interest in egocentric human video as a scalable, substantially lower-cost, and more diverse alternative for embodied model pretraining. However, its effectiveness compared to teleoperated real-robot data remains underexplored. To address this question, we conduct a systematic study comparing egocentric human video and teleoperated real-robot trajectories as pretraining data sources for embodied foundation models, under fixed post-training and validation protocols. Surprisingly, we find that egocentric data, when processed through a carefully designed filtering and labeling pipeline, is not merely a viable substitute for model pretraining but can lead to superior performance. With the same amount of pretraining data, models pretrained on egocentric data achieve a 24% lower validation loss on real-robot action prediction, as well as 52.5% and 90% higher success rates on in-distribution and out-of-distribution real-robot task execution, respectively. This finding verifies a scalable paradigm for embodied foundation models: pretrain on egocentric human video to learn diverse world representations, then adapt with a small amount of labeled real-robot data for action-space alignment. We hope this study encourages broader exploration of egocentric data and offers guidance for data quality assessment before costly robot data collection.
Human behavior is among the most scalable sources of data for learning physical intelligence, yet how to effectively leverage it for dexterous manipulation remains unclear. While prior work demonstrates human to robot transfer in constrained settings, it is unclear whether large scale human data can support fine grained, high degree of freedom dexterous manipulation. We present EgoScale, a human to dexterous manipulation transfer framework built on large scale egocentric human data. We train a Vision Language Action (VLA) model on over 20,854 hours of action labeled egocentric human video, more than 20 times larger than prior efforts, and uncover a log linear scaling law between human data scale and validation loss. This validation loss strongly correlates with downstream real robot performance, establishing large scale human data as a predictable supervision source. Beyond scale, we introduce a simple two stage transfer recipe: large scale human pretraining followed by lightweight aligned human robot mid training. This enables strong long horizon dexterous manipulation and one shot task adaptation with minimal robot supervision. Our final policy improves average success rate by 54% over a no pretraining baseline using a 22 DoF dexterous robotic hand, and transfers effectively to robots with lower DoF hands, indicating that large scale human motion provides a reusable, embodiment agnostic motor prior.
Ruijie Zheng, Dantong Niu, Yuqi Xie +12
NVIDIA · University of California, Berkeley · University of Maryland
Egocentric human video offers a scalable data source for dexterous manipulation, yet using it to train humanoid robots presents two challenges: (1) an embodiment gap, as human hands differ structurally from robot end-effectors and low-cost egocentric recordings lack the torso kinematics required by conventional retargeting; and (2) heterogeneous data quality, including noisy hand-pose tracking and weakly aligned text annotations. We introduce IronMind, a vision-language-action (VLA) model that uses egocentric human video and heterogeneous robot data to pretrain policies for humanoid dexterous manipulation. To bridge the embodiment gap, IronMind bypasses explicit body-retargeting by using a camera-space action representation, the native reference space of egocentric video, and semantically aligning robot and human action dimensions. Across total pretraining budgets from 250 to 10,000 hours, validation loss decreases approximately log-linearly with data scale. Larger pretraining budgets also improve out-of-distribution real-robot manipulation after post-training: across six challenging tasks with unseen objects, affordances, and reasoning prompts, the 10,000-hour model achieves a 55.0% success rate, compared with at most 11.7% for every pretraining budget up to 5,000 hours and 5.0% without pretraining. At the same pretraining budget, the camera-space action representation also outperforms the torso-frame baseline. Together, these findings support pretraining with a camera-space action representation on large-scale human egocentric data as a scalable foundation for humanoid robot manipulation.