Underwater C3-JEPA: An Object-Centric Cross-View World Model for ROV Salvage

arXiv 2026 43.3 method, application

TLDR

Underwater C3-JEPA is a control-conditioned object-centric multi-view latent world model for ROV salvage, predicting task-object state for MPC and behavior-agent training.

评分理由

The paper's strengths include a novel object-centric cross-view predictive architecture and real-world validation, with low annotation cost via weak binding. Weaknesses are limited visibility into quantitative comparisons, baseline details, and the fact that the latent-state focus may not fully justify the 'world model' framing.

Read-first 评分解释

综合优先阅读分 43.3，由主题、引用、图谱、方法、可复现性和近期性等信号加权得到。原始总分保留为 47。

方法质量 15%

100

检查可见的摘要与分析字段，寻找实验、数据集、基线、指标和局限性等方法证据。命中信号：基线、基准、评估、实验、指标、结果、验证

近期性 5%

100

使用温和的时间衰减，让近期论文更容易浮现，同时保留较早基础工作的价值。年份：2026

主题相关性 25%

67.1

使用现有 LLM 关键词相关性评分，并归一化到 0-100。关键词：world model、world simulator、generative world model、interactive world model、video world model、world dynamics prediction、model-based reinforcement learning world model

可复现性 15%

检查链接和可见文本中的论文、代码、数据集、工件与仓库信号。论文：有；代码：无；数据：无；命中信号：代码

图谱声望 15%

5.7

基于可用 OpenAlex 引用关系构建局部引用图，并计算 PageRank 风格的图谱声望。局部 PageRank：0.0016

引用影响力 15%

使用 OpenAlex 形态的引用元数据作为文献关注度信号，并与论文本身质量分开处理。引用数：0

引用速度 10%

引用速度按发表年限估算年均引用，降低旧论文天然占优的偏差。年均引用：0.00

研究版图角色

前沿论文方法锚点

排序敏感性

稳定性：volatile；排名波动范围：751。

关键词评分

world model

world dynamics prediction

interactive world model

model-based reinforcement learning world model

video world model

world simulator

generative world model

深度分析

创新点

Object-centric multi-view predictive world model (Underwater C3-JEPA) for near-field heavy-load underwater ROV salvage that predicts task-object state evolution in latent space without contact sensors.
Cross-view, control-conditioned, context-extended architecture that encodes multi-camera observations into task-object and context tokens, fuses cross-camera evidence via held-out-view attention, and directly predicts future states conditioned on control.
Weak binding mechanism to anchor target and gripper at low annotation cost.
SIGReg to sharpen geometric representation.
Predictive interface supporting model-predictive-control candidate evaluation and imagined-rollout behavior-agent training.

方法

The model encodes synchronized multi-view RGB observations and vehicle control signals into task-object and context tokens, fuses cross-camera evidence using held-out-view attention, and directly predicts future latent states conditioned on control. Weak binding anchors target and gripper at low annotation cost, while SIGReg sharpens geometric representations. Evaluation compares downstream probe transfer against a reconstruction-free latent baseline and validates on real underwater video using withheld-camera object-state recovery and persistence benchmarks.

关键结果

The learned representation transfers substantially more task-relevant information to downstream probes than a reconstruction-free latent baseline while keeping the predictor lightweight. Validation on real underwater video shows the same architecture recovers a withheld camera's object state and stays ahead of persistence, indicating the recipe transfers beyond simulation.

资源

论文