Underwater C3-JEPA: An Object-Centric Cross-View World Model for ROV Salvage
Yuncong Yang, Jinlong Li, Yulong Xue, Feng Wu, Chunwen Zhang, Lei Qiao, Xuyang Wang
cs.RO · cs.AI
Quick verdict
Object-centric, control-conditioned JEPA world model for ROV salvage; careful ablations and open-loop field tests, but sim-only closed loop, no code release.
Summary
The paper presents Underwater C3-JEPA, an object-centric world model for near-field ROV salvage that predicts, without contact sensors, how target and gripper states evolve under vehicle commands, from synchronized multi-view RGB and control signals. Frozen DINOv3 features feed a temporal slot encoder; weak binding anchors stabilize target and gripper identity; held-out-view cross-attention fuses cameras; a 1.33M-parameter control-conditioned JEPA predictor rolls out latent states for MPC candidate scoring and imagined-rollout agent training. In ROS-Gazebo simulation, binding guidance raises task-object group Dice to about 0.90 versus 0.09-0.12 for unguided slots at a modest feature-MSE cost, and held-out-view attention beats weighted-sum fusion on all five axes of an equal-budget ladder. On a 229-recording diagnostic, terminal relative-motion error falls from 39.1 cm (persistence) to 25.7 cm; matched-endpoint MPC ranking picks the best candidate in 53.3% of 90 groups versus 25% chance. Against a reconstruction-free LeWM baseline trained on identical data, frozen probes show roughly halved joint-angle and yaw error. Two field deployments on uncalibrated co-mounted cameras recover a withheld view at 81-88% of achievable range and beat persistence by 23-31% at 5 s. Limitations: single-run ablation arms, a non-recording-disjoint held-out evaluation, agent gains measured by the training critic, simplified underwater optics, open-loop field tests in relative units, and no released code.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 5
● low 1
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Weak-binding object-centric slots preserve target and gripper identity (group Dice about 0.90/0.88 vs 0.09-0.12 for unguided slots) at a small reconstruction cost.
- Held-out-view cross-attention lets a single fused token retain evidence for a camera it never saw (withheld-view Dice 0.2206 to 0.7604 for the UUV, 0.2034 to 0.8695 for the gripper).
- A 1.33M-parameter control-conditioned JEPA predictor beats persistence (terminal error 39.1 to 25.7 cm) and ranks MPC candidates above chance (53.3% vs 25% over 90 groups).
- Object-centric reconstruction guidance transfers roughly twice the task information of a reconstruction-free LeWM baseline under identical frozen probes.
- The same recipe trains from raw field recordings, recovering a withheld camera at 81-88% of achievable range and beating persistence by 23-31% at 5 s, in open-loop relative units.
Strengths
Tiered evidence with negative controls
Sec. IV-B defines a five-level evidence ladder and states that an improvement at one level is not evidence for a later one; Sec. V-B uses zeroed/shuffled-control interventions, defends the persistence baseline with citations, and reports where persistence remains competitive (latent-MSE parity at 10 s on AutoGrasp).
Honest reporting of negative results
The gripper decoder Dice is reported worse under held-out fusion (0.891 vs 0.902/0.903); field rollout mask quality fails to improve (0.159 vs 0.184, CI -0.053 to +0.002) and is explicitly separated from token fidelity; the Note to Practitioners states closed-loop sea success is unmeasured.
Isolated attribution of the object-centric contribution
Mechanism-level fusion evidence
Deployment-scale accounting
Weaknesses
No closed-loop or task-success evidence on hardware
Central ablations rest on single runs and a leaky split
Field fusion test is weak
Simplified underwater imaging
Reproducibility gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This IEEE submission (arXiv:2609.30214, cs.RO/cs.AI) presents Underwater C3-JEPA, an object-centric world model for near-field heavy-load ROV salvage. Three claims are bundled: (1) an object-centric perception domain (OCPD) with guided cross-view fusion (GO-CVF), where weak-binding anchors fix target and gripper identity, held-out-view attention fuses the cameras, and free slots carry context; (2) a control-conditioned JEPA predictor (1.33M parameters) whose rollouts support MPC candidate ranking and imagined-rollout behavior-agent training; (3) transfer of the same recipe to real pool footage. The authors are appropriately explicit that the deliverable is a predictive state interface, not an autonomous salvage system: closed-loop success exists only in simulation ('an interface demonstration, not a powered success-rate study', Sec. VI), and all field results are open-loop in relative units.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Has this robot actually performed autonomous underwater salvage in the ocean?
No. Closed-loop salvage behavior is demonstrated only in ROS-Gazebo simulation, coupled to a conventional controller, and the authors call it an interface demonstration rather than a success-rate study. The two real deployments were in a test pool (60 m x 20 m, 10 m deep) and were open-loop replays reported in relative units; closed-loop sea trials are stated as future work.
Does C3-JEPA require extra sensors or hardware on the ROV?
No. It trains on the video and the control/telemetry a work-class ROV already logs, with no contact sensors. The predictor is small: 1.33M parameters, 6.5M FLOPs per window, about 1.97 ms per forward pass and 76 MB memory on an A800 GPU, though an embedded implementation is left to future work.
How is C3-JEPA different from Dreamer, V-JEPA 2, or other world models?
It adds object-centric structure: two weakly-bound slots fix target and gripper identity, held-out-view cross-attention fuses the ROV's cameras while withholding one view during training, and free slots carry context. Compared against a reconstruction-free latent world model (LeWM) trained on identical data, frozen probes recover about twice the task information (joint-angle error 4.90 vs 9.81 degrees; yaw 6.50 vs 13.82 degrees).
Is the code or dataset for C3-JEPA publicly available?
Not publicly. The paper states data and code are available from the corresponding authors upon reasonable request. Public components exist (UUV Simulator, DINOv3, VideoSAUR), but the 705-recording release manifests, the weak-anchor annotations, and the evaluation protocols in supplementary sections S2-S6 are not released.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-9 are referenced but not analyzable from the extracted text; numeric claims rely on body text and table transcriptions.
- Supplementary S2-S6 (training details, evaluation protocols, Table S4) is not included; several results (e.g., the 15 s direction-cosine split 0.42 vs 0.23) defer to it.
- Equations and table formatting are partially garbled by extraction; loss weights and architecture hyperparameters were read from Table I.
- arXiv v1 (Sep 2026), submitted to IEEE but not yet peer-reviewed at review time.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.