AD-L-JEPA: Self-Supervised Spatial World Models with Joint Embedding Predictive Architecture for Autonomous Driving with LiDAR Data

arXiv 25.1 2025 49.6 method, application

TLDR

AD-L-JEPA is a self-supervised JEPA-based pre-training framework for LiDAR 3D object detection that predicts Bird's-Eye-View embeddings, achieving state-of-the-art results with reduced compute.

Reasoning

Strengths include a novel non-generative, non-contrastive approach with strong empirical results on multiple large-scale datasets and significant compute savings. Weaknesses are its narrow focus on LiDAR object detection without general world modeling capabilities, and no discussion of limitations.

Read-first score

Read-first score 49.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 6.

Recency 8%

86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=True; dataset=False; markers=code,dataset,github

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=dataset,experiment,result

Topical relevance 42%

8.6

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Field roles

FrontierMethodology anchorReproducibility anchor

Rank sensitivity

Stability: volatile; rank range: 644.

Keyword Scores

world model

generative world model

world dynamics prediction

world simulator

interactive world model

video world model

model-based reinforcement learning world model

Deep Analysis

Innovations

First JEPA-based pre-training method for autonomous driving with LiDAR data.
Predicts Bird's-Eye-View embeddings instead of explicitly generating masked regions.
Uses explicit variance regularization to avoid representation collapse without manual contrastive pairs.
Achieves better quality, faster training (1.9x-2.7x GPU hours reduction), and lower GPU memory (2.8x-4x reduction) compared to Occupancy-MAE.

Methodology

AD-L-JEPA is a self-supervised pre-training framework based on a joint embedding predictive architecture (JEPA) for LiDAR 3D object detection. It predicts Bird's-Eye-View embeddings from masked LiDAR data and employs explicit variance regularization to prevent representation collapse, avoiding both generative and contrastive approaches. The model is pre-trained on unlabeled LiDAR data and evaluated on downstream detection tasks using KITTI3D, Waymo, and ONCE datasets.

Key Results

On the ONCE dataset, pre-training on 100K frames yields a 1.61 mAP gain, outperforming all other methods pre-trained on up to 500K frames; pre-training on 500K frames yields a 2.98 mAP gain, outperforming all methods pre-trained on up to 1M frames. Additionally, AD-L-JEPA reduces GPU hours by 1.9x-2.7x and GPU memory by 2.8x-4x compared to Occupancy-MAE.

Resources

Paper Code