In four real-robot manipulation tasks, every tested future-vision setup scored above an action-only training baseline, while none of the three latent-action setups did. The authors report the result in an Oct. 15 arXiv preprint, using 20 trials per task and a score that gives partial credit.
Training routes samples to action or one auxiliary target
Vision-language-action (VLA) policies use camera images and language instructions to produce robot movements. The study asks whether a policy must spend most of its training on predicting those movements. The researchers compare direct action prediction with two targets derived from demonstrations: future vision (FV), which predicts features of a later camera view, and latent action (LA), which compresses a visual transition into tokens describing how the scene changes.
A sample-level router assigns each training-sample occurrence to either direct action prediction or one auxiliary target. At P50, an expected half of sample occurrences use the direct-action loss and half use the selected auxiliary loss. For FV-routed examples, the model predicts later visual features while receiving the demonstrated action as an input. At deployment, the policy takes the current view, instruction and robot state and outputs actions; it receives no future view or auxiliary target.
Future-vision targets led on four robot tasks
The strongest FV version, using Wan2.2 visual features, scored 156/160 when picking objects out of a basket, 112/160 when placing them back, 108/160 when rearranging cups and 148/160 when rearranging objects. The action-only baseline scored 116/160, 92/160, 80/160 and 124/160, respectively. Each task was tested over 20 trials, with up to eight points per trial; these totals measure partial task progress, not fully successful episodes.
All seven policy configurations used the same 30,000-update training budget on one DOS W1 robot arm. The training set contained 203 demonstrations across the four tasks. The paper counts repeated uses of a demonstration as additional sample exposure, but says a matched sample budget does not establish equal computation or training time. The small task set and single hardware platform leave open how the result would transfer to other robots and jobs.
Sources
Study and appendix; arXiv record, posted Oct. 15, 2026. Illustrative A-Lab image, not the robot used in the study: Wikimedia Commons file by the A-Lab study authors, licensed under CC BY 4.0.
.webp){kind=link}