Robot training has gained a cheaper route. Galbot, Peking University, the Institute of Automation at the Chinese Academy of Sciences and Tsinghua University feed an ordinary first-person human video into a temporary memory module. The robot reads the scene and the job first, then uses its existing action ability in the deployment environment.
The framework is WAM-TTT, or World-Action Model Test-Time Training. The practical question is whether a robot trained to move cups in a lab can enter a user home, with different light and a different tabletop, and start with little new robot data.
Watching video is not copying a hand
“At test time, only unlabeled human videos are required to adapt the memory, while the pretrained WAM remains frozen.”
Only a lightweight fast-weight memory is updated from unlabeled human video. The base world-action model stays frozen, unlike pipelines that estimate hand pose, retarget motion or collect another batch of teleoperated robot trajectories.
Nine tasks, three robot forms
The team tested a Unitree G1 humanoid, a Galbot two-finger gripper and a Galbot dexterous hand on 9 tasks, including pouring water, delivering a drink, clearing a table, flipping steak and stacking a pyramid. Each task-environment pair ran 25 trials and was scored by task progress.
- WAM-TTT averaged 46.2% progress in new household environments.
- A frozen WAM backbone with LDA reached 32.5%.
- WAM-ICL, using human video directly as context, reached 7.1%.
- A reproduced EgoScale baseline reached 15.0%.
Table Bussing reached 100.0% and Swap Place 66.7%, while Stamp Paper fell to 8.3%, below LDA's 33.3%. The paper points to tighter geometric pose requirements and household disturbances.
The savings are in robot data
Embodied AI is expensive because actions must be collected across desks, objects and lighting. WAM-TTT shifts part of that deployment data from robot trajectories to ordinary human videos. In an ablation, 100 human videos plus 100 robot trajectories reached 74.1% average success; a hand-pose and motion-retargeting pipeline averaged only 28.9% completion across 4 tasks.
The next metric to watch is cross-scene degradation. In Deliver Drink disturbance tests, WAM-TTT reached 66.0% under lighting change and 56.0% under spatial change; WAM-ICL reached 12.0% and 20.0%. It is not yet a robot that can do all housework from a casual video, but it makes post-deployment adaptation measurable.
Sources: QbitAI; arXiv:2607.06988 was used to verify model structure, task settings, metrics and limitations; CocoLoop, Hugging Face Papers were used to verify the paper entry and public indexing status.