U4D: Uncertainty-Aware 4D World Modeling from LiDAR Sequences

Conference: CVPR 2026
arXiv: 2512.02982
Code: None
Area: Autonomous Driving
Keywords: LiDAR Generation, Uncertainty Modeling, Diffusion Models, 4D World Model, Spatio-Temporal Consistency

TL;DR

Ours proposes U4D, the first uncertainty-aware 4D LiDAR world modeling framework. It adopts a "hard-to-easy" two-stage diffusion generation strategy, first reconstructing high-uncertainty regions and then conditionally completing the entire scene. A MoST module is designed to adaptively fuse spatio-temporal features to ensure temporal consistency.

Background & Motivation

LiDAR data acquisition bottleneck: Collecting large-scale, diverse, and annotated LiDAR data is extremely costly and labor-intensive. Generative LiDAR modeling has become a crucial path for data augmentation and pre-training.

Uniform assumption of existing methods: Methods like LiDARGen, LiDM, and R2DM treat all spatial regions equally during generation, ignoring the non-uniform distribution of semantic difficulty in real-world scenes.

Existence of uncertainty regions: Long-range sparse areas, occluded object boundaries, small-scale structures, and semantically ambiguous regions naturally exhibit high uncertainty in LiDAR observations. Uniform generation leads to geometric artifacts and temporal instability in these areas.

Human-like cognitive inspiration: Humans parse ambiguous regions before understanding the global context when perceiving a scene. U4D draws on this idea by first generating uncertainty regions as structural anchors and then completing the rest of the scene.

Insufficient temporal consistency: Existing methods mainly focus on spatial reconstruction and lack modeling of temporal coherence between frames, resulting in unnatural object motion in generated sequences.

Downstream task-driven demand: Safety-critical applications like autonomous driving require generated data to truly improve the robustness and calibration reliability of perception models, rather than just pursuing visual fidelity.

Method

Overall Architecture

U4D adopts a two-stage "hard-to-easy" generation paradigm:

• Stage 1: Uncertainty Region Modeling — A pre-trained LiDAR segmentation model (RangeNet++) is used to estimate point-wise uncertainty maps (Shannon entropy). Top-K high-entropy points are selected to form a sparse uncertainty point cloud. After converting to range-view representation, an unconditional diffusion model reconstructs high-fidelity uncertainty regions.
• Stage 2: Uncertainty-Conditioned Completion — Using the uncertainty regions generated in the first stage as conditional input, a conditional diffusion model completes the full LiDAR frames, ensuring global structural consistency.
• The two stages share a unified latent scene representation, allowing global context information to refine local uncertainties.
• A MoST Spatio-Temporal Fusion Module is embedded within the diffusion backbones of both stages, adaptively fusing features from the spatial branch (intra-frame geometry) and the temporal branch (inter-frame dynamics) via gating to maintain temporal coherence while ensuring single-frame fidelity.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400, 'subGraphTitleMargin': {'top': 8, 'bottom': 16}}}}%%
flowchart TD
    A["LiDAR Sequence Input"] --> B["Uncertainty Measurement & Representation<br/>RangeNet++ estimates point-wise Shannon entropy<br/>→ Select Top-K high-entropy points → range-view"]
    B --> C["Uncertainty Region Diffusion Model<br/>Unconditional diffusion reconstructs high-entropy regions + validity mask"]
    C -->|"Generated uncertainty regions as conditional prior"| D["Uncertainty-Conditioned Completion<br/>Conditional diffusion completes full frames using uncertainty regions as anchors"]
    D --> E["4D LiDAR Sequence Output"]
    subgraph MoST["MoST Spatio-Temporal Fusion Module (Embedded in two-stage diffusion backbones)"]
        direction TB
        M1["Spatial Branch 1×3×3<br/>Intra-frame geometry"] --> M3["Gated Adaptive Fusion<br/>+ Noise / CV Regularization"]
        M2["Temporal Branch 3×1×1<br/>Inter-frame dynamics"] --> M3
    end
    C -.Spatio-temporal feature fusion.-> MoST
    D -.Spatio-temporal feature fusion.-> MoST

Key Designs

1. Uncertainty Measurement and Representation

• Calculate Shannon entropy for each point: \(H(\mathbf{p}) = -\sum_{c=1}^{C} D_c(\mathbf{p}) \log D_c(\mathbf{p})\)
• nuScenes retains Top-20% high-entropy points; SemanticKITTI retains Top-5%.
• Sparse uncertainty point clouds are projected into range images \(\mathbf{x}_0^u \in \mathbb{R}^{H \times W \times 2}\) (depth + intensity), accompanied by a binary mask \(\mathbf{m}^u\).

2. Uncertainty Region Diffusion Model

• An unconditional diffusion model \(\epsilon_\theta^u\) learns the generation distribution of uncertainty regions in range-view.
• It employs a standard DDPM forward process, reconstructing the spatial validity mask during reverse denoising.

3. Uncertainty-Conditioned Completion

• A conditional diffusion model \(\epsilon_\theta^c\) learns \(p(\mathbf{x}_0 | \mathbf{x}_0^u)\).
• Noisy input \(\mathbf{x}_t\) and the uncertainty prior \(\mathbf{x}_0^u\) are concatenated along the feature dimension, enabling the network to utilize both global and local cues for denoising.

4. MoST (Mixture of Spatio-Temporal) Module

• Intermediate features \(\mathbf{F}_i \in \mathbb{R}^{C_i \times L \times H_i \times W_i}\) are decomposed into a spatial branch (\(1 \times 3 \times 3\) convolution for intra-frame geometry) and a temporal branch (\(3 \times 1 \times 1\) convolution for inter-frame dynamics).
• Outputs from both branches are concatenated and passed through an MLP shared embedding, then adaptively fused via a MoE-style gating mechanism: \((α_i^s, α_i^t) = \text{Softmax}(\mathbf{F}_i^{\text{share}} \cdot \mathbf{W}_i^g + \mathbb{I}(\chi \cdot \sigma(\mathbf{F}_i^{\text{share}} \cdot \mathbf{W}_i^z)))\)
• Gaussian noise perturbations are added to the gating during training to avoid deterministic overfitting.
• The spatial branch dominates input/output layers (geometric details), while the temporal branch dominates intermediate layers (motion dynamics).

Loss & Training

• Uncertainty stage loss: \(\mathcal{L}_u = \mathbb{E}[\|\epsilon^u - \epsilon_\theta^u(\mathbf{x}_t^u, t)\|_2^2] + \lambda \mathcal{L}_{\text{mask}}(\mathbf{m}^u, \mathbf{m}^p)\), where \(\mathcal{L}_{\text{mask}}\) is binary cross-entropy.
• Conditioned completion loss: \(\mathcal{L}_c = \mathbb{E}[\|\epsilon^c - \epsilon_\theta^c(\mathbf{x}_t, t, \mathbf{x}_0^u)\|_2^2]\)
• Gating regularization: \(\mathcal{L}_{\text{reg},i} = \frac{\text{Var}(\alpha_i^s)}{(\mathbb{E}[\alpha_i^s])^2} + \frac{\text{Var}(\alpha_i^t)}{(\mathbb{E}[\alpha_i^t])^2}\), preventing the gate from excessively favoring a single modality.
• Two stages are trained separately: 1M steps for the first stage, 0.5M steps for the second stage; batch size 8, sequence length 6.
• AdamW optimizer, learning rate \(1 \times 10^{-4}\), cosine annealing + 10K steps warmup; EMA decay rate 0.995.
• 4 × NVIDIA RTX 4090, FP16 mixed precision training.

Key Experimental Results

Main Results

Table 1: nuScenes Scene-level Generation Fidelity

Method	FRD ↓	FPD ↓	JSD ↓	MMD(×10⁻⁴) ↓
LiDARGen (ECCV'22)	549.18	22.80	0.04	0.76
R2DM (ICRA'24)	253.80	14.35	0.03	0.48
UniScene (CVPR'25)	-	976.47	0.32	13.61
Ours (U4D)	223.96	12.90	0.03	0.53

Table 3: nuScenes Temporal Consistency (TTCE/CTC)

Method	TTCE-3 ↓	TTCE-4 ↓	CTC-1 ↓	CTC-3 ↓
UniScene (CVPR'25)	2.74	3.69	0.90	3.64
LiDARCrafter (AAAI'26)	2.65	3.56	1.12	3.02
Ours (U4D)	2.63	3.51	0.97	2.98

Table 4: Downstream Semantic Segmentation mIoU(%)

Method	1% Labels	10% Labels	50% Labels
Sup.-only	58.3	71.0	75.1
R2DM	64.1	73.0	75.9
Ours (U4D)	65.3	73.7	76.4

Ablation Study

Uncertainty Region Selection Strategy (Table 6)

Strategy	FRD ↓	FPD ↓	ECE(%) ↓
No Uncertainty	235.91	14.03	3.98
Random Sampling	235.23	13.21	4.35
Confidence Sampling	228.24	13.04	3.02
Entropy Sampling	223.96	12.90	2.72

MoST Fusion Strategy (Table 7)

Fusion Method	FRD ↓	FPD ↓	JSD ↓
Cascade (No Parallel)	536.23	23.34	0.63
Additive Fusion	242.81	13.42	0.28
Concatenation Fusion	242.43	12.51	0.03
Adaptive Fusion	223.96	12.90	0.03

Key Findings

• Entropy-based uncertainty selection significantly outperforms random selection and unconditional generation, with ECE dropping from 3.98% to 2.72%.
• Adaptive fusion yields a significant FRD improvement (~20 points) over simple addition/concatenation, validating the effectiveness of dynamic gating.
• U4D generated data provides the largest gain (+7.0 mIoU) for downstream segmentation at low annotation ratios (1%), indicating that uncertainty-aware generated data is especially valuable for data-scarce scenarios.
• MoST exhibits different spatio-temporal activation patterns across network depths: spatial-dominant in shallow/deep layers and temporal-dominant in intermediate layers.

Highlights & Insights

• Pioneering Uncertainty-Driven LiDAR Generation: Fed the uncertainty output of perception models back into the generation process, creating a "Perception → Uncertainty → Generation → Perception Enhancement" loop.
• "Hard-to-Easy" Generation Philosophy: Conquering semantically ambiguous "hard" regions first and using them as anchors to complete "easy" regions is a strategy with broad applicability.
• Exquisite MoST Module Design: Gating mechanism + Noise regularization + Coefficient of variation (CV) regularization provides a triple guarantee for balanced spatio-temporal feature fusion.
• Downstream calibration experiments (ECE metric) prove that generated data can not only improve accuracy but also enhance model confidence calibration.

Limitations & Future Work

• Inference speed (8.9s/frame) is higher than single-frame methods (R2DM 3.5s); efficiency still needs improvement.
• Uncertainty estimation depends on the quality of the pre-trained segmentation model; biases in the segmentation model may propagate to the generation process.
• Validated only on nuScenes and SemanticKITTI; more sensor configurations and indoor scenes have not been tested.
• The two-stage training pipeline (1.5M steps total) is computationally expensive; an end-to-end solution might be more efficient.
• Information loss in range-view representation for self-occlusion and long-range areas is not fully discussed.

Related Work & Insights

• LiDAR Generation: LiDARGen (ECCV'22 Score-based) → R2DM (ICRA'24 Diffusion) → LiDARCrafter (AAAI'26 Auto-regressive temporal) → U4D (Uncertainty-aware).
• Uncertainty Modeling: SalsaNext (Bayesian inference), Calib3D (Depth-aware calibration) — U4D is the first to introduce uncertainty into a generative framework.
• Spatio-Temporal Modeling: ViDAR (Image-to-LiDAR prediction), Cascaded spatio-temporal in video diffusion — MoST proposes parallel decomposition + adaptive gating as an alternative to cascade.

Rating

• Novelty: ⭐⭐⭐⭐ (Uncertainty-driven "hard-to-easy" generation is a fresh perspective)
• Experimental Thoroughness: ⭐⭐⭐⭐ (Dual datasets, multiple metrics, complete ablation, downstream task validation)
• Writing Quality: ⭐⭐⭐⭐ (Clear structure, excellent diagrams, well-articulated motivation)
• Value: ⭐⭐⭐⭐ (The uncertainty+generation paradigm has practical significance for autonomous driving simulation data augmentation)

Related Papers

• [AAAI 2026] LiDARCrafter: Dynamic 4D World Modeling from LiDAR Sequences
• [CVPR 2026] GenieDrive: Towards Physics-Aware Driving World Model with 4D Occupancy Guided Video Generation
• [CVPR 2026] FoSS: Modeling Long-Range Dependencies and Multimodal Uncertainty in Trajectory Prediction via Fourier–State Space Integration
• [CVPR 2026] Bezier Degradation Modeling for LiDAR-based Human Motion Capture
• [CVPR 2026] Points-to-3D: Structure-Aware 3D Generation with Point Cloud Priors