WAM4D: Fast 4D World Action Model via Spatial Register Tokens

A new 4D world action model called WAM4D uses spatial register tokens to give robots better spatial awareness while keeping inference fast, according to a paper posted to arXiv on June 12, 2026 [1][2]. World action models, or WAMs, jointly predict future visual observations and the executable actions a robot should take. Most existing WAMs work in 2D video or latent spaces, where visually plausible rollouts lack the 3D spatial constraints and occluded contact geometry needed for precise manipulation [1][2]. Geometric foundation models can recover dense 3D structure and motion from visual data, but forcing a WAM to predict a full 4D representation introduces costly geometric decoding that slows causal action generation [2]. To address that trade-off, the researchers designed WAM4D with lightweight spatial register tokens. During training, these tokens serve as future-depth readouts, transferring pretrained geometric priors into a causal video-action transformer. For inference, the register branch is removed entirely, preserving speed [1][2]. The team also introduced causal mixture attention for the Mixture-of-Transformers backbone, which defines modality-specific visibility among video, action, and geometry tokens to prevent non-causal shortcuts [2]. Experiments on the RoboTwin 2.0 dataset and real-world manipulation tasks showed that WAM4D improves spatial consistency and achieves competitive action prediction while maintaining efficient inference [1][2]. The paper appeared on arXiv, the open-access e-print repository that hosts preprints across physics, computer science, and related fields [6]. As of November 2024, arXiv was receiving about 24,000 submissions per month and had surpassed two million total articles by the end of 2021 [6]. The work was shared through arXivLabs, a framework that lets community collaborators build experimental tools directly on the platform [4]. arXivLabs projects range from bibliographic explorers to code finders and recommender systems, all operating under guidelines that require partners to uphold openness, community, excellence, and user data privacy [4][5]. The arXivLabs program is currently pausing new proposals while the development team focuses on modernizing and migrating arXiv’s systems to the cloud [3].