A robot policy trained with more human-video examples succeeded more often in a small real-world test, rising from 63.3% to 85% as the study's annotated video pool grew from 1,000 to 5,000 hours. The result comes from a new arXiv preprint, and it does not mean the robot learned to act from video alone: the system still used hundreds of hours of robot data and task demonstrations.
Turning video into action-like supervision
Human videos contain abundant examples of people manipulating objects, but they do not record the commands a robot needs to execute. The authors' LAPT-Scale method addresses that gap by learning a continuous, low-dimensional “latent action” from the visual change between two frames. The representation is a training target that summarizes an observed transition; it is not itself a motor command that can be sent to a robot.
The paper describes three training stages. First, a latent-action model is trained on 23,300 hours of egocentric video, then used to annotate a cleaned 5,100-hour subset for policy pretraining. Second, the policy is aligned to a target robot using 250 hours of robot data. Third, it is adapted to specific tasks with robot demonstrations. For the three original real-robot tasks, the paper says that final stage used 50 demonstrations per task.
This builds on earlier work such as LAPA, which also learned action-like representations from videos before mapping them to robot actions. LAPT-Scale's reported changes include a continuous rather than quantized latent space, perceptual and semantic training signals, and optical-flow alignment intended to make the representation track movement and object interaction rather than mostly static pixels.
More video, same robot-data setup
The central scaling comparison varied the amount of annotated human video used in the first policy-training stage: 1,000, 3,000 or 5,000 hours. The researchers say they kept the policy architecture, robot data and evaluation protocol fixed. On an AgiBot G1, the policy performed three tasks involving a mouse pad, headphones and a cup, with 20 trials per task. Average success increased from 63.3% at 1,000 hours to 80% at 3,000 and 85% at 5,000 hours.
A control experiment provides context for the result. The policy without human-video pretraining scored 55%. In a compute-matched control, the authors extended robot-alignment training, and success was 53.3%. These comparisons support the authors' case that the video-derived supervision helped in this setup. They are experiments from the same paper, not an independent replication.
A promising signal, not a general robot score
The study shows a relationship between more annotated human video and better results in one controlled pipeline. It does not isolate video from robot data: robot alignment and task-specific demonstrations remained part of training. Nor does a rise across three tabletop tasks establish that the policy will work across different robots, homes or industrial settings.
The authors also report stronger performance than a comparison policy on four harder tasks under tested changes in position, lighting and background. But their own paper cautions that real-robot coverage is limited to one target platform and a small task set, with only 20 trials per task and condition. Larger studies across robots and tasks, plus independent replication, would be needed to test how far the result generalizes.
For now, LAPT-Scale is best read as evidence that action-like signals inferred from human video can complement robot training data. The practical question is whether that benefit survives when the platform, task distribution and data pipeline change.
Sources
- Xiang et al., “LAPT-Scale: Scaling Embodied Policies with Latent Action Pre-Training on Human Videos,” full paper (submitted October 9, 2026; Figure 1 reproduced under CC BY 4.0).
- arXiv record and abstract for LAPT-Scale.
- Ye et al., “Latent Action Pretraining from Videos” (LAPA), earlier related work.