RepWAM: World Action Modeling with Representation Visual-Action Tokenizers

A research team has introduced RepWAM, a representation-centric world action model that departs from conventional video tokenizers by operating in a semantic visual-action latent space, according to a paper posted to arXiv on 11 June 2026 [1][2]. Existing world action models typically inherit reconstruction-oriented video tokenizers from pretrained video generation models. While those tokenizers preserve visual fidelity, the authors argue that pixel reconstruction alone offers limited guidance for learning the instruction-following dynamics that link future prediction with robot control [1][2]. To address this gap, RepWAM trains a representation visual-action tokenizer that maps visual inputs into aligned visual and latent action tokens [1][2]. The model is then pretrained to jointly model future visual states and the latent actions that connect them under language instructions, before being adapted to real robot trajectories for closed-loop manipulation [1][2]. Experiments on real-world manipulation tasks and simulation benchmarks show that RepWAM delivers strong performance across diverse manipulation settings. Ablation studies highlight the value of semantic visual-action tokenization over reconstruction-oriented alternatives [1][2]. The authors state that these results establish representation visual-action tokenization as a promising foundation for world action models and a step toward generalist robot policies [1][2]. Code and weights will be made available on GitHub [2]. The paper appeared on arXiv, the open-access e-print repository that hosts preprints across physics, mathematics, computer science, and related fields [6]. arXiv, which began in 1991, surpassed two million articles by the end of 2021 and receives roughly 24,000 submissions per month as of late 2024 [6]. The RepWAM abstract page includes links to community tools developed under arXivLabs, a framework launched in 2020 that allows third-party collaborators to build experimental features directly on the site [5]. These tools include the Bibliographic Explorer for navigating citation trees, the CORE Recommender for surfacing related open-access papers, and integrations with Hugging Face and Papers with Code [4][5]. arXivLabs partners must adhere to values of openness, community, excellence, and user data privacy, and they receive only minimal, anonymized data for feature functionality [5].