Envision4D: Envisioning Visual Futures via Feed-forward 4D Gaussian Splatting for Autonomous Driving

A new self-supervised framework called Envision4D can forecast future driving scenes without requiring pre-set camera poses, according to a paper posted to the arXiv preprint server [1]. The method uses a feed-forward design to extrapolate dynamic environments frame by frame [2]. The paper, submitted on June 9, 2026, describes a system built to address a persistent problem in autonomous-driving research: existing feed-forward models are tuned for interpolation between known frames and produce ghosting artifacts when asked to predict farther into the future [2]. Envision4D tackles this by introducing a Future Pose Prediction module that infers upcoming camera parameters through an iterative denoising process, removing the need for fixed pose priors [2]. To handle the non-linear motion of vehicles, pedestrians, and cyclists, the authors added In-layer Temporal Attention and a technique called Conditioned Motion Lifting. The latter recasts the uncertain extrapolation task as a set of relational mappings, which the paper states leads to more stable predictions [2]. A Progressive Training Strategy is also used to keep unsupervised motion learning from drifting as errors accumulate over longer time horizons [2]. The framework is fully self-supervised, meaning it learns directly from raw video without annotated ground-truth poses or depth maps [2]. In experiments reported in the preprint, Envision4D achieved state-of-the-art results on future view synthesis, outperforming prior methods by a significant margin [2]. arXiv, where the work appears, is an open-access repository that hosts electronic preprints across physics, computer science, mathematics, and related fields [6]. Papers on the site are moderated but not peer-reviewed before posting [6]. As of late 2024, the repository was receiving roughly 24,000 new articles per month and had surpassed two million total submissions by the end of 2021 [6]. The platform also supports community-built tools through its arXivLabs program, which lets third-party developers offer features such as citation explorers and code-finding widgets directly on abstract pages [5]. No external funding or institutional affiliations were disclosed in the preprint abstract, and the research has not yet appeared in a peer-reviewed journal [1].