Latent Action Pretraining Through World Modeling

A new self-supervised framework called LAWM can pretrain robot imitation models on unlabeled video, sidestepping the costly teleoperation data that dominates current vision-language-action models, its authors report [1]. The framework, detailed in a paper revised in June 2026, learns latent action representations by modeling how the visual world changes between frames, a technique that allows it to use recordings of robots or even humans handling everyday objects [1]. Current state-of-the-art vision-language-action models, or VLAs, such as OpenVLA and π₀, rely on large-scale datasets of manually labeled actions collected through teleoperation [1]. VLAs are multimodal foundation models that take an image and a text instruction and directly output low-level robot actions [7]. The concept was pioneered by Google DeepMind with RT-2 in July 2023 [7]. More recent approaches, including LAPA and villa-X, introduced latent action representations for unsupervised pretraining, but their large model sizes have made real-world deployment difficult [1]. LAWM is designed to be model-agnostic and efficient enough for practical settings [1]. The authors report that it outperforms models pretrained with ground-truth robot actions and other similar pretraining methods on the LIBERO benchmark and in real-world tests [1]. A language model benchmark is a standardized test with a dataset and evaluation metrics used to compare models on tasks such as language understanding and generation [6]. The work arrives as AI research continues to push toward systems that can perceive, reason, and act. Artificial intelligence, as a field founded in 1956, has moved through cycles of optimism and so-called AI winters before the current boom driven by deep learning and the transformer architecture [5]. Self-supervised pretraining, which LAWM employs, echoes earlier breakthroughs in natural language processing. BERT, introduced by Google researchers in 2018, learned contextual representations of text through masked token prediction and became a ubiquitous baseline [4]. The paper was submitted by Bahey Tharwat on September 22, 2025, with a second version posted on June 13, 2026 [1]. The first submission weighed 5,756 KB and the second 7,170 KB [1].