Vision-language-action (VLA) policies are usually trained to predict actions, since that is what robots ultimately execute. A new arXiv paper, "What Should Robot Policies Learn to Predict? Rethinking Supervision Targets in Vision-Language-Action Learning," challenges this assumption. The abstract notes that auxiliary objectives—such as latent-action (LA) and future-vision (FV)—may offer additional supervision signals.
The source text, however, is cut off mid-sentence, so the paper's full argument and findings are not available from this abstract alone. The title and opening suggest a rethinking of what VLA models should be trained to predict, but the specific conclusions remain unclear.
Given the incomplete abstract, readers should treat this as an early indication of a research direction rather than a settled result. The paper appears to be asking a fundamental question about the design of robot learning objectives, but the answer is not yet visible from the provided text.