4D Scene Understanding

We introduce OVOW, the first training-free system that reconstructs instance-level, simulation-ready 4D mesh scenes from a single monocular video. Recent 4D reconstruction achieves impressive rendering quality, but its outputs (\eg, implicit fields, Gaussian primitives, or point clouds) lack the watertight topology, instance separation, and standardized physical interfaces required by physics simulators and embodied AI. OVOW closes this gap with a four-stage pipeline: a vision-language model discovers, labels, and motion-classifies all instances; category-aware reconstruction yields per-instance meshes for rigid objects and topology-consistent mesh sequences for deformable ones; an iterative render-match-optimize procedure recovers metric scale and 6-DoF pose trajectories; and physics-grounded assembly enforces ground contact and inter-object support. Crucially, we model all motion, rigid and non-rigid, through direct vertex deformation without category-specific priors or skeleton rigging, producing watertight mesh scenes ready for downstream physics simulation and editing. We further establish the first benchmark for structured Video-to-4D evaluation, with metrics for geometric correctness, instance separation, and physical plausibility beyond visual fidelity; the same pipeline doubles as a scalable engine for synthesizing paired video-to-4D simulation data for future 4D world models and embodied AI. Across two synthetic benchmarks (static and 4D), OVOW attains the best overall layout and geometry accuracy and the lowest photometric and semantic error among all baselines, and on monocular video runs one to two orders of magnitude faster than the baselines, while downstream physics simulation confirms its physical stability.

Language-guided dynamic scene understanding requires identifying one physical entity across time, metric 3D space, and multiple viewpoints, a setting not captured by current multimodal reasoning or tracking paradigms. We introduce 4DVLT, an offline calibrated multi-view task that unifies grounding with metric 3D and synchronized 2D tracking through a worldline binding identity, motion, visibility, and observations. We construct Instruct-4D, with 129.4K QA pairs, 64.7K targets, 851 scenes, and 9 reasoning types, and propose 4DTrack for query-conditioned inference over a 4D state graph. 4DTrack reaches 55.18 WQS, exceeding the strongest adapted VLT baseline by 33.01 points while improving grounding and trajectory quality. The project page is available at https://github.com/mikubaka88/4DVLT.

Despite recent advances, Vision Language Models (VLMs) still struggle to grasp the dynamics of the world. We note that the ability to reason about a 4D scene, challenging in itself, is further complicated by two factors. First, VLMs observe motion indirectly via its projection onto 2D images. Second, existing datasets fail to disentangle object and camera motion. To address these challenges, we present a QA generation pipeline that focuses on motion-related scene understanding. We take particular care of the entanglement of camera and object motion by casting tracking in both the traditional way and in a novel, fixed reference system, dubbed True-Motion Tracking, which provides an intuitive description of motion. From this pipeline, we generate a large-scale training dataset of 400K samples, 4DP-QA (4D Perception QA), and a 2.2K-sample benchmark, 4DP-QA-Bench. Training existing models on our dataset yields performance improvements on an external benchmark, validating the effectiveness of our method.

Constructing faithful 4D worlds from LiDAR-acquired sequences is crucial for embodied AI, yet current generative frameworks apply uniform modeling capacity across all spatial regions. This ignores that perceptual difficulty varies dramatically within a single scan: distant surfaces, occluded boundaries, and small-scale objects carry far higher uncertainty than well-observed structures. We present U4D, a new framework that explicitly leverages spatial uncertainty to guide LiDAR scene generation in a "hard-to-easy" schedule. U4D derives per-point uncertainty maps via Shannon Entropy from a pretrained segmentor, then applies an unconditional diffusion stage to synthesize high-entropy areas with precise geometry, followed by a conditional completion stage that fills in the remaining regions using these structures as priors. A MoST (Mixture of Spatio-Temporal) block further maintains cross-frame coherence by dynamically balancing spatial detail and temporal continuity. Extensive experiments on nuScenes and SemanticKITTI demonstrate state-of-the-art scene fidelity, temporal consistency, and downstream performance.

Recent feed-forward 3D gaussian splatting methods have made dramatic progress on individual aspects of 3D scene reconstruction, but no existing method jointly addresses dynamic content, multi-view input, and unknown camera poses in a single feed-forward pass. Methods that handle dynamics either require accurate camera poses or accept only monocular input; pose-free multi-view methods address only static scenes; and per-scene optimization methods bridge some of these gaps but at minutes-to-hours cost per scene. We introduce NoPo4D, the first feed-forward system that addresses this empty quadrant. Building on a pretrained geometry backbone and recent 4D Gaussian frameworks, NoPo4D introduces a velocity decomposition that splits Gaussian motion into per-pixel image-plane shifts and depth changes, allowing direct supervision from pseudo ground-truth optical flow on the 2D component. This sidesteps both the differentiable rendering that couples prior posed methods to pose accuracy and the 3D motion ground truth that prior pose-free methods require. The system is rounded out by a bidirectional motion encoder for cross-view and cross-frame feature aggregation, and view-dependent opacity that mitigates cross-view and cross-timestep Gaussian misalignments. On four multi-view dynamic benchmarks, NoPo4D consistently outperforms prior feed-forward baselines, and with an optional post-optimization stage surpasses per-scene optimization methods, while running orders of magnitude faster.

Recent large multimodal models (LMMs) have become increasingly capable on image and video understanding, yet still struggle to sustain 4D continuous spatiotemporal dynamic reasoning. To study this capability gap, we formulate trajectory-grounded multi-turn spatiotemporal dialogue, a new task in which a model must answer spatiotemporal queries while returning structured 3D target trajectories over an entire short clip or a specified segment of a longer clip, and introduce Track4D-Bench, a benchmark with 526 clip-level dialogue samples spanning 23.5k frames and 7.5k object annotations, for training and evaluation. Building on this task, we propose LMM-Track4D, which combines RTGE (Ray--Time Geometry Encoding), a dedicated streaming state token TRK for long-horizon dynamic propagation, and an Object-Slot Kinematic, Residual-Anchor (OSK-RA) decoder for stable 4-step 3D state estimation under occlusion and viewpoint variation. Experiments on Track4D-Bench show consistent improvements over strong baselines, suggesting that explicit dynamic state modeling is a useful design principle for eliciting 4D dynamic reasoning in LMMs. Our code and dataset will be publicly available at https://github.com/mikubaka88/LMM-Track4D.

Dynamic spatial reasoning from monocular video is essential for bridging visual intelligence and the physical world, yet remains challenging for vision-language models (VLMs). Prior approaches either verbalize spatial-temporal reasoning entirely as text, which is inherently verbose and imprecise for complex dynamics, or rely on external geometric modules that increase inference complexity without fostering intrinsic model capability. In this paper, we present 4DThinker, the first framework that enables VLMs to "think with 4D" through dynamic latent mental imagery, i.e., internally simulating how scenes evolve within the continuous hidden space. Specifically, we first introduce a scalable, annotation-free data generation pipeline that synthesizes 4D reasoning data from raw videos. We then propose Dynamic-Imagery Fine-Tuning (DIFT), which jointly supervises textual tokens and 4D latents to ground the model in dynamic visual semantics. Building on this, 4D Reinforcement Learning (4DRL) further tackles complex reasoning tasks via outcome-based rewards, restricting policy gradients to text tokens to ensure stable optimization. Extensive experiments across multiple dynamic spatial reasoning benchmarks demonstrate that 4DThinker consistently outperforms strong baselines and offers a new perspective toward 4D reasoning in VLMs. Our code is available at https://github.com/zhangquanchen/4DThinker.

Dense 3D reconstruction and tracking of dynamic scenes from monocular video remains an important open challenge in computer vision. Progress in this area has been constrained by the scarcity of high-quality datasets with dense, complete, and accurate geometric annotations. To address this limitation, we introduce Syn4D, a multiview synthetic dataset of dynamic scenes that includes ground-truth camera motion, depth maps, dense tracking, and parametric human pose annotations. A key feature of Syn4D is the ability to unproject any pixel into 3D to any time and to any camera. We conduct extensive evaluations across multiple downstream tasks to demonstrate the utility and effectiveness of the proposed dataset, including 4D scene reconstruction, 3D point tracking, geometry-aware camera retargeting, and human pose estimation. The experimental results highlight Syn4D's potential to facilitate research in dynamic scene understanding and spatiotemporal modeling.

Understanding human activities and their surrounding environments typically relies on visual perception, yet cameras pose persistent challenges in privacy, safety, energy efficiency, and scalability. We explore an alternative: 4D perception without vision. Its goal is to reconstruct human motion and 3D scene layouts purely from everyday wearable sensors. For this we introduce IMU-to-4D, a framework that repurposes large language models for non-visual spatiotemporal understanding of human-scene dynamics. IMU-to-4D uses data from a few inertial sensors from earbuds, watches, or smartphones and predicts detailed 4D human motion together with coarse scene structure. Experiments across diverse human-scene datasets show that IMU-to-4D yields more coherent and temporally stable results than SoTA cascaded pipelines, suggesting wearable motion sensors alone can support rich 4D understanding.

We present Vista4D, a robust and flexible video reshooting framework that grounds the input video and target cameras in a 4D point cloud. Specifically, given an input video, our method re-synthesizes the scene with the same dynamics from a different camera trajectory and viewpoint. Existing video reshooting methods often struggle with depth estimation artifacts of real-world dynamic videos, while also failing to preserve content appearance and failing to maintain precise camera control for challenging new trajectories. We build a 4D-grounded point cloud representation with static pixel segmentation and 4D reconstruction to explicitly preserve seen content and provide rich camera signals, and we train with reconstructed multiview dynamic data for robustness against point cloud artifacts during real-world inference. Our results demonstrate improved 4D consistency, camera control, and visual quality compared to state-of-the-art baselines under a variety of videos and camera paths. Moreover, our method generalizes to real-world applications such as dynamic scene expansion and 4D scene recomposition. See our project page for results, code, and models: https://eyeline-labs.github.io/Vista4D

4D point cloud videos capture rich spatial and temporal dynamics of scenes which possess unique values in various 4D understanding tasks. However, most existing methods work in the spatiotemporal domain where the underlying geometric characteristics of 4D point cloud videos are hard to capture, leading to degraded representation learning and understanding of 4D point cloud videos. We address the above challenge from a complementary spectral perspective. By transforming 4D point cloud videos into graph spectral signals, we can decompose them into multiple frequency bands each of which captures distinct geometric structures of point cloud videos. Our spectral analysis reveals that the decomposed low-frequency signals capture more coarse shapes while high-frequency signals encode more fine-grained geometry details. Building on these observations, we design Spatio-Temporal-Spectral Mixer (STS-Mixer), a unified framework that mixes spatial, temporal, and spectral representations of point cloud videos. STS-Mixer integrates multi-band delineated spectral signals with spatiotemporal information to capture rich geometries and temporal dynamics, while enabling fine-grained and holistic understanding of 4D point cloud videos. Extensive experiments show that STS-Mixer achieves superior performance consistently across multiple widely adopted benchmarks on both 3D action recognition and 4D semantic segmentation tasks. Code and models are available at https://github.com/Vegetebird/STS-Mixer.

Understanding dynamic 4D environments through natural language queries requires not only accurate scene reconstruction but also robust semantic grounding across space, time, and viewpoints. While recent methods using neural representations have advanced 4D reconstruction, they remain limited in contextual reasoning, especially for complex semantics such as interactions, temporal actions, and spatial relations. A key challenge lies in transforming noisy, view-dependent predictions into globally consistent 4D interpretations. We introduce PanopticQuery, a framework for unified query-time reasoning in 4D scenes. Our approach builds on 4D Gaussian Splatting for high-fidelity dynamic reconstruction and introduces a multi-view semantic consensus mechanism that grounds natural language queries by aggregating 2D semantic predictions across multiple views and time frames. This process filters inconsistent outputs, enforces geometric consistency, and lifts 2D semantics into structured 4D groundings via neural field optimization. To support evaluation, we present Panoptic-L4D, a new benchmark for language-based querying in dynamic scenes. Experiments demonstrate that PanopticQuery sets a new state of the art on complex language queries, effectively handling attributes, actions, spatial relationships, and multi-object interactions. A video demonstration is available in the supplementary materials.

Volumetric video seeks to model dynamic scenes as temporally coherent 4D representations. While recent Gaussian-based approaches achieve impressive rendering fidelity, they primarily emphasize appearance but are largely agnostic to instance-level structure, limiting stable tracking and semantic reasoning in highly dynamic scenarios. In this paper, we present Director, a unified spatio-temporal Gaussian representation that jointly models human performance, high-fidelity rendering, and instance-level semantics. Our key insight is that embedding instance-consistent semantics naturally complements 4D modeling, enabling more accurate scene decomposition while supporting robust dynamic scene understanding. To this end, we leverage temporally aligned instance masks and sentence embeddings derived from Multimodal Large Language Models to supervise the learnable semantic features of each Gaussian via two MLP decoders, enabling language-aligned 4D representations and enforcing identity consistency over time. To enhance temporal stability, we bridge 2D optical flow with 4D Gaussians and finetune their motions, yielding reliable initialization and reducing drift. For the training, we further introduce a geometry-aware SDF constraints, along with regularization terms that enforces surface continuity, enhancing temporal coherence in dynamic foreground modeling. Experiments demonstrate that Director achieves temporally coherent 4D reconstructions while simultaneously enabling instance segmentation and open-vocabulary querying.

Spatiotemporal reasoning is a fundamental capability for artificial intelligence (AI) in soft tissue surgery, paving the way for intelligent assistive systems and autonomous robotics. While 2D vision-language models show increasing promise at understanding surgical video, the spatial complexity of surgical scenes suggests that reasoning systems may benefit from explicit 4D representations. Here, we propose a framework for equipping surgical agents with spatiotemporal tools based on an explicit 4D representation, enabling AI systems to ground their natural language reasoning in both time and 3D space. Leveraging models for point tracking, depth, and segmentation, we develop a coherent 4D model with spatiotemporally consistent tool and tissue semantics. A Multimodal Large Language Model (MLLM) then acts as an agent on tools derived from the explicit 4D representation (e.g., trajectories) without any fine-tuning. We evaluate our method on a new dataset of 134 clinically relevant questions and find that the combination of a general purpose reasoning backbone and our 4D representation significantly improves spatiotemporal understanding and allows for 4D grounding. We demonstrate that spatiotemporal intelligence can be "assembled" from 2D MLLMs and 3D computer vision models without additional training. Code, data, and examples are available at https://tum-ai.github.io/surg4d/

We present THFM, a unified video foundation model for human-centric perception that jointly addresses dense tasks (depth, normals, segmentation, dense pose) and sparse tasks (2d/3d keypoint estimation) within a single architecture. THFM is derived from a pretrained text-to-video diffusion model, repurposed as a single-forward-pass perception model and augmented with learnable tokens for sparse predictions. Modulated by the text prompt, our single unified model is capable of performing various perception tasks. Crucially, our model is on-par or surpassing state-of-the-art specialized models on a variety of benchmarks despite being trained exclusively on synthetic data (i.e.~without training on real-world or benchmark specific data). We further highlight intriguing emergent properties of our model, which we attribute to the underlying diffusion-based video representation. For example, our model trained on videos with a single human in the scene generalizes to multiple humans and other object classes such as anthropomorphic characters and animals -- a capability that hasn't been demonstrated in the past.

Feedforward reconstruction is crucial for autonomous driving applications, where rapid scene reconstruction enables efficient utilization of large-scale driving datasets in closed-loop simulation and other downstream tasks, eliminating the need for time-consuming per-scene optimization. We present StreetForward, a pose-free and tracker-free feedforward framework for dynamic street reconstruction. Building upon the alternating attention mechanism from Visual Geometry Grounded Transformer (VGGT), we propose a simple yet effective temporal mask attention module that captures dynamic motion information from image sequences and produces motion-aware latent representations. Static content and dynamic instances are represented uniformly with 3D Gaussian Splatting, and are optimized jointly by cross-frame rendering with spatio-temporal consistency, allowing the model to infer per-pixel velocities and produce high-fidelity novel views at new poses and times. We train and evaluate our model on the Waymo Open Dataset, demonstrating superior performance on novel view synthesis and depth estimation compared to existing methods. Furthermore, zero-shot inference on CARLA and other datasets validates the generalization capability of our approach. More visualizations are available on our project page: https://streetforward.github.io.

We present Inst4DGS, an instance-decomposed 4D Gaussian Splatting (4DGS) approach with long-horizon per-Gaussian trajectories. While dynamic 4DGS has advanced rapidly, instance-decomposed 4DGS remains underexplored, largely due to the difficulty of associating inconsistent instance labels across independently segmented multi-view videos. We address this challenge by introducing per-video label-permutation latents that learn cross-video instance matches through a differentiable Sinkhorn layer, enabling direct multi-view supervision with consistent identity preservation. This explicit label alignment yields sharp decision boundaries and temporally stable identities without identity drift. To further improve efficiency, we propose instance-decomposed motion scaffolds that provide low-dimensional motion bases per object for long-horizon trajectory optimization. Experiments on Panoptic Studio and Neural3DV show that Inst4DGS jointly supports tracking and instance decomposition while achieving state-of-the-art rendering and segmentation quality. On the Panoptic Studio dataset, Inst4DGS improves PSNR from 26.10 to 28.36, and instance mIoU from 0.6310 to 0.9129, over the strongest baseline.

Current 4D representations decouple geometry, motion, and semantics: reconstruction methods discard interpretable motion structure; language-grounded methods attach semantics after motion is learned, blind to how objects move; and motion-aware methods encode dynamics as opaque per-point residuals without object-level organization. We propose 4D Synchronized Fields, a 4D Gaussian representation that learns object-factored motion in-loop during reconstruction and synchronizes language to the resulting kinematics through a per-object conditioned field. Each Gaussian trajectory is decomposed into shared object motion plus an implicit residual, and a kinematic-conditioned ridge map predicts temporal semantic variation, yielding a single representation in which reconstruction, motion, and semantics are structurally coupled and enabling open-vocabulary temporal queries that retrieve both objects and moments. On HyperNeRF, 4D Synchronized Fields achieves 28.52 dB mean PSNR, the highest among all language-grounded and motion-aware baselines, within 1.5 dB of reconstruction-only methods. On targeted temporal-state retrieval, the kinematic-conditioned field attains 0.884 mean accuracy, 0.815 mean vIoU, and 0.733 mean tIoU, surpassing 4D LangSplat (0.620, 0.433, and 0.439 respectively) and LangSplat (0.415, 0.304, and 0.262). Ablation confirms that kinematic conditioning is the primary driver, accounting for +0.45 tIoU over a static-embedding-only baseline. 4D Synchronized Fields is the only method that jointly exposes interpretable motion primitives and temporally grounded language fields from a single trained representation. Code will be released.

Spatio-temporal scene graphs provide a principled representation for modeling evolving object interactions, yet existing methods remain fundamentally frame-centric: they reason only about currently visible objects, discard entities upon occlusion, and operate in 2D. To address this, we first introduce ActionGenome4D, a dataset that upgrades Action Genome videos into 4D scenes via feed-forward 3D reconstruction, world-frame oriented bounding boxes for every object involved in actions, and dense relationship annotations including for objects that are temporarily unobserved due to occlusion or camera motion. Building on this data, we formalize World Scene Graph Generation (WSGG), the task of constructing a world scene graph at each timestamp that encompasses all interacting objects in the scene, both observed and unobserved. We then propose three complementary methods, each exploring a different inductive bias for reasoning about unobserved objects: PWG (Persistent World Graph), which implements object permanence via a zero-order feature buffer; MWAE (Masked World Auto-Encoder), which reframes unobserved-object reasoning as masked completion with cross-view associative retrieval; and 4DST (4D Scene Transformer), which replaces the static buffer with differentiable per-object temporal attention enriched by 3D motion and camera-pose features. We further design and evaluate the performance of strong open-source Vision-Language Models on the WSGG task via a suite of Graph RAG-based approaches, establishing baselines for unlocalized relationship prediction. WSGG thus advances video scene understanding toward world-centric, temporally persistent, and interpretable scene reasoning.

Humans inhabit a physical 4D world where geometric structure and semantic content evolve over time, constituting a dynamic 4D reality (spatial with temporal dimension). While current Multimodal Large Language Models (MLLMs) excel in static visual understanding, can they also be adept at "thinking in dynamics", i.e., perceive, track and reason about spatio-temporal dynamics in evolving scenes? To systematically assess their spatio-temporal reasoning and localized dynamics perception capabilities, we introduce Dyn-Bench, a large-scale benchmark built from diverse real-world and synthetic video datasets, enabling robust and scalable evaluation of spatio-temporal understanding. Through multi-stage filtering from massive 2D and 4D data sources, Dyn-Bench provides a high-quality collection of dynamic scenes, comprising 1k videos, 7k visual question answering (VQA) pairs, and 3k dynamic object grounding pairs. We probe general, spatial and region-level MLLMs to express how they think in dynamics both linguistically and visually, and find that existing models cannot simultaneously maintain strong performance in both spatio-temporal reasoning and dynamic object grounding, often producing inconsistent interpretations of motion and interaction. Notably, conventional prompting strategies (e.g., chain-of-thought or caption-based hints) provide limited improvement, whereas structured integration approaches, including Mask-Guided Fusion and Spatio-Temporal Textual Cognitive Map (ST-TCM), significantly enhance MLLMs' dynamics perception and spatio-temporal reasoning in the physical 4D world. Code and benchmark are available at https://dyn-bench.github.io/.

Predicting scene dynamics from visual observations is challenging. Existing methods capture dynamics only within observed boundaries failing to extrapolate far beyond the training sequence. Node-RF (Neural ODE-based NeRF) overcomes this limitation by integrating Neural Ordinary Differential Equations (NODEs) with dynamic Neural Radiance Fields (NeRFs), enabling a continuous-time, spatiotemporal representation that generalizes beyond observed trajectories at constant memory cost. From visual input, Node-RF learns an implicit scene state that evolves over time via an ODE solver, propagating feature embeddings via differential calculus. A NeRF-based renderer interprets calculated embeddings to synthesize arbitrary views for long-range extrapolation. Training on multiple motion sequences with shared dynamics allows for generalization to unseen conditions. Our experiments demonstrate that Node-RF can characterize abstract system behavior without explicit model to identify critical points for future predictions.

Effective environment modeling is the foundation for autonomous driving, underpinning tasks from perception to planning. However, current paradigms often inadequately consider the feedback of ego motion to the observation, which leads to an incomplete understanding of the driving process and consequently limits the planning capability. To address this issue, we introduce a novel ego-scene interactive modeling paradigm. Inspired by human recognition, the paradigm represents ego-scene interaction as the scene flow relative to the ego-vehicle. This conceptualization allows for modeling ego-motion feedback within a feature learning pattern, advantageously utilizing existing log-replay datasets rather than relying on scenario simulations. We specifically propose FlowAD, a general flow-based framework for autonomous driving. Within it, an ego-guided scene partition first constructs basic flow units to quantify scene flow. The ego-vehicle's forward direction and steering velocity directly shape the partition, which reflects ego motion. Then, based on flow units, spatial and temporal flow predictions are performed to model dynamics of scene flow, encompassing both spatial displacement and temporal variation. The final task-aware enhancement exploits learned spatio-temporal flow dynamics to benefit diverse tasks through object and region-level strategies. We also propose a novel Frames before Correct Planning (FCP) metric to assess the scene understanding capability. Experiments in both open and closed-loop evaluations demonstrate FlowAD's generality and effectiveness across perception, end-to-end planning, and VLM analysis. Notably, FlowAD reduces 19% collision rate over SparseDrive with FCP improvements of 1.39 frames (60%) on nuScenes, and achieves an impressive driving score of 51.77 on Bench2Drive, proving the superiority. Code, model, and configurations will be released here.

Egocentric video understanding is inherently complex due to the dynamic 4D nature of the environment, where camera motion and object displacements necessitate a continuous re-evaluation of spatial relations. In this work, we target a suite of under-explored egocentric 4D reasoning tasks, including fixture interaction counting, viewpoint-relative fixture location, object movement itinerary tracking, and stationary object localization, that require fundamentally different cognitive operations: spatial anchoring, temporal tracking, and duration reasoning. We observe that these structural differences make task-agnostic approaches insufficient: generic Chain-of-Thought methods lack task-appropriate reasoning primitives, and uniform reinforcement learning actively destabilizes performance on spatial tasks. To address this, we propose EgoReasoner, a two-stage framework that aligns both the reasoning scaffold and the reward signal to each task's cognitive structure. In the first stage, Task-Adaptive Thinking Templates guide the synthesis of structured CoT traces that teach the model to reason adaptively across task types via supervised fine-tuning. In the second stage, task-aware reward functions verify entity grounding, temporal alignment, and task-adaptive logical consistency, selectively strengthening each reasoning pathway via reinforcement fine-tuning with GRPO. Our 3B-parameter model, trained on only 16K samples, achieves 37.5% average accuracy on the challenging HD-EPIC benchmark, surpassing Qwen2.5-VL-7B (25.7%) by over 10 points.

Humans are born with vision-based 4D spatial-temporal intelligence, which enables us to perceive and reason about the evolution of 3D space over time from purely visual inputs. Despite its importance, this capability remains a significant bottleneck for current multimodal large language models (MLLMs). To tackle this challenge, we introduce MLLM-4D, a comprehensive framework designed to bridge the gaps in training data curation and model post-training for spatiotemporal understanding and reasoning. On the data front, we develop a cost-efficient data curation pipeline that repurposes existing stereo video datasets into high-quality 4D spatiotemporal instructional data. This results in the MLLM4D-2M and MLLM4D-R1-30k datasets for Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT), alongside MLLM4D-Bench for comprehensive evaluation. Regarding model training, our post-training strategy establishes a foundational 4D understanding via SFT and further catalyzes 4D reasoning capabilities by employing Group Relative Policy Optimization (GRPO) with specialized Spatiotemporal Chain of Thought (ST-CoT) prompting and Spatiotemporal reward functions (ST-reward) without involving the modification of architecture. Extensive experiments demonstrate that MLLM-4D achieves state-of-the-art spatial-temporal understanding and reasoning capabilities from purely 2D RGB inputs. Project page: https://github.com/GVCLab/MLLM-4D.

Capturing 4D spatiotemporal scene structure is crucial for the safe and reliable operation of robots in dynamic environments. However, existing approaches typically address only part of the problem: they either provide coarse geometric tracking via bounding boxes or detailed 3D occupancy estimates that lack explicit temporal association and instance-level reasoning. In this work, we present Latent Gaussian Splatting (LaGS) for 4D Panoptic Occupancy Tracking (4D-POT). We revisit the underlying representation and model 3D features as a sparse set of feature-bearing Gaussians. These act as dynamic, volume-oriented keypoints that enable spatially continuous, distance-weighted aggregation of multi-view features before being splatted into a voxel grid for decoding. This point-centric formulation enables flexible, data-dependent receptive fields and long-range spatial interactions that are difficult to capture with local and dense voxel-based operators. A hierarchical Gaussian representation further enables multi-scale reasoning by combining global context from coarse super-points with fine-grained detail from higher-resolution streams. Extensive experiments on Occ3D nuScenes and Waymo demonstrate state-of-the-art performance for 4D-POT. We provide code and models at https://lags.cs.uni-freiburg.de/.

Point cloud video understanding is critical for robotics as it accurately encodes motion and scene interaction. We recognize that 4D datasets are far scarcer than 3D ones, which hampers the scalability of self-supervised 4D models. A promising alternative is to transfer 3D pre-trained models to 4D perception tasks. However, rigorous empirical analysis reveals two critical limitations that impede transfer capability: overfitting and the modality gap. To overcome these challenges, we develop a novel "Align then Adapt" (PointATA) paradigm that decomposes parameter-efficient transfer learning into two sequential stages. Optimal-transport theory is employed to quantify the distributional discrepancy between 3D and 4D datasets, enabling our proposed point align embedder to be trained in Stage 1 to alleviate the underlying modality gap. To mitigate overfitting, an efficient point-video adapter and a spatial-context encoder are integrated into the frozen 3D backbone to enhance temporal modeling capacity in Stage 2. Notably, with the above engineering-oriented designs, PointATA enables a pre-trained 3D model without temporal knowledge to reason about dynamic video content at a smaller parameter cost compared to previous work. Extensive experiments show that PointATA can match or even outperform strong full fine-tuning models, whilst enjoying the advantage of parameter efficiency, e.g. 97.21 % accuracy on 3D action recognition, $+8.7 \%$ on 4 D action segmentation, and 84.06% on 4D semantic segmentation.

Despite rapid progress, embodied navigation in dynamic and unstructured urban environments remains brittle. Most existing approaches directly map monocular visual inputs to actions through end-to-end pixel-to-action training, assuming that accurate spatiotemporal (4D) scene understanding will emerge implicitly. While appealing, this paradigm requires large amounts of pixel-to-action supervision that are difficult to obtain. This challenge is amplified in dynamic, unstructured settings, where robust navigation requires precise 4D scene modeling. To address these limitations, we present Stereo4DWalker, a 4D-aware embodied navigation model that leverages stereo inputs and explicitly builds structured 4D representations of geometry and motion. These 4D structures are integrated into the navigation transformer through simple yet effective 4D-conditioned attention layers. To support scalable training, we curate a large-scale stereo navigation dataset with automatically annotated actions from Internet stereo videos. Our experiments show that Stereo4DWalker surpasses state-of-the-art performance using only 1.5% of the training data, highlighting the effectiveness of explicit 4D visual modeling for data-efficient and robust urban navigation.