Ego4WAM: What Matters When Scaling Egocentric Human Data for Robot Learning?

Which properties of egocentric human data and which training-stage usage strategies most improve downstream closed-loop robot manipulation when scaling human data under a fixed world-action model framework?

Egocentric human data is cheaper and more diverse than robot teleoperation, but it varies widely in human-robot alignment, behavioral coverage, and action-label quality. Existing large-scale studies treat total data duration as the main scaling axis and often rely on open-loop human-action prediction, so it is unclear whether gains come from better alignment, broader task coverage, video-only supervision, or simply more data.

Prior work used large-scale egocentric datasets such as Ego4D, EgoDex, and EgoScale; retargeted human actions to pretrain manipulation priors; and applied world-action models such as EgoWAM, OpenWAM, and FastWAM. These approaches typically aggregate scaling curves or report human-action prediction error, which obscures which data properties drive robot performance and how video-only versus action-labeled data should be staged.

The authors fix a world-action model backbone with a video branch and an action branch, define a unified 32D bimanual action space separating shared end-effector/gripper components from robot-specific joints, and organize egocentric data into a pyramid: robot-aligned demonstrations, action-validated multi-task video-action data, and large-scale video-only data. They use three stages—video-only world-model pre-training, optional joint world-action mid-training, and downstream robot post-training—and compare base versus joint shared attention. Evaluation is closed-loop on a real bimanual Piper robot and the RoboDojo benchmark.

On real-robot Place into the Basket, robot-only training achieved 60% in-domain success but only 10% object-OOD and 0% scene-OOD; adding object-OOD human demonstrations raised object-OOD success to 60%, and adding scene-OOD human demonstrations raised scene-OOD success to 20%. On Fold Cloth, 100 robot demonstrations plus 500 aligned human demonstrations recovered 80% in-domain, 60% object-OOD, and 80% scene-OOD success; with only 20 robot demonstrations, aligned human data alone was not enough, but after multi-task mid-training the same 20 robot demonstrations plus aligned human data reached 80% in-domain, 80% object-OOD, and 60% scene-OOD. On RoboDojo, no human pre-training or mid-training scored 6.39 average score and 3.15% success rate; video-only pre-training alone reached 14.13/9.45%, and switching to joint attention improved it to 20.29/14.35%. This Ego4WAM configuration used 0.1K robot hours and 15K egocentric hours, compared with Xiaomi-1 at 20.07/13.93%, GalaxeaVLA-0.5 at 20.23/14.88%, OpenWAM-alpha at 17.18/11.92%, and GPT-6-Astra at 28.97/22.48%. Data-duration scaling from 1K to 4K hours improved RoboDojo Generalization from 5.0 to 7.8 and Long-Horizon from 12.0 to 16.2; task-diversity scaling from 500 to 6K tasks improved Generalization to 7.0 and Long-Horizon to 16.2 but reduced Precision.

The paper is a preprint and has not been peer-reviewed. Real-robot quantitative results are limited to a few tasks on a single bimanual Piper platform, and broader real-robot tasks in Section 4.4 are shown without reported metrics. The benchmark is RoboDojo, not real-world multi-embodiment deployment. The model performs poorly on the Open category (0.55 score, 0.50% success rate) and is not designed for long-term memory or open VLM instruction following. Action labels depend on hand- and camera-pose estimation, and many samples are filtered out or used only as video. Long-tail task diversity can degrade precision. The final strong result uses 15K hours of video pre-training, implying substantial compute, and code/checkpoints are promised but not yet released.