LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

A new robot-control architecture called LaWAM uses compact latent visual predictions instead of full video generation to guide manipulation tasks, achieving state-of-the-art success rates while cutting inference latency by up to 24 times compared to pixel-space world-action models, according to a preprint posted on arXiv [1]. The system, formally named Latent World Action Model, was described in a paper submitted to the arXiv preprint repository on 14 June 2026 [1]. arXiv, which began operating in 1991, is an open-access repository that hosts electronic preprints across physics, computer science, and other fields without peer review [6]. The repository passed two million articles by the end of 2021 and now receives roughly 24,000 submissions per month [6]. LaWAM addresses a known gap in Vision-Language-Action models, which leverage large-scale vision-language pretraining for semantic robot control but often lack explicit foresight into how actions change a scene [2]. World-Action Models, or WAMs, condition policies on predicted futures, yet existing implementations typically rely on computationally expensive video generation that carries substantial pixel-level redundancy [2]. At the core of LaWAM is a latent-action-conditioned component called the Latent World Model, or LaWM [2]. The researchers trained a latent action model inside the latent space of a pretrained vision foundation model and repurposed its forward decoder to predict future observation features [2]. Action generation is then conditioned on these predicted latent visual subgoals rather than on reconstructed future video frames [1]. The paper reports that LaWAM achieved a 98.6 percent success rate on the LIBERO benchmark and a 91.22 percent success rate on the RoboTwin benchmark, along with competitive results on real-world manipulation tasks [1]. Inference runs in 187 milliseconds per action-chunk prediction, which the authors state is up to 24 times lower in wall-clock latency than comparable pixel-space WAMs [2]. The preprint appears on arXiv with links to several community-built tools available through the arXivLabs framework, including the Bibliographic Explorer and the CORE Recommender [4][5]. arXivLabs, formalized in 2020, allows third-party collaborators to develop experimental features that appear on article record pages under a framework that enforces user-data privacy and aligns with arXiv’s values of openness and community [4].