acceptodds
Under review as a conference paper at ICLR 2027

DepState-JEPA: Learning Depth-Grounded State Representation for Visual Perception and Manipulation

Abstract

Predictive representations for robot manipulation should connect scene geometry with the state changes induced by actions. Our intuition is that understanding what an action will change starts with understanding where things are: complementary RGB and depth observations ground a common scene representation from which to infer those changes. Guided by this intuition, we introduce **DepState-JEPA**, a dual-stream RGB–Depth joint-embedding predictive architecture that connects spatial understanding with action-conditioned prediction through a shared latent state, **DepState**. Built on V-JEPA 2.1, the framework uses depth both as a complementary observation stream and to ground both modalities in metric 3D space, aggregating shared information into DepState while retaining modality-private features for decoding. Ray-conditioned readouts predict local features without target depth; after an action-conditioned transition, the same decoders predict future RGB and depth features using the predicted state and fixed initial private features. Robot-state readouts provide recursive proprioceptive inputs; object-trajectory supervision and calibrated forward kinematics shape state learning through motion targets and joint–TCP consistency. Comparisons show stronger RGB-encoder transfer after RGB-D adaptation, feature completion, conditioned visual forecasting, and manipulation success. Within-model ablations identify contributions from depth content, metric coordinates, and future-state matching, with additional control gains from FK regularization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Visual Perception, JEPA, Depth, Embodied Manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_YXmzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2(+1)
Read review
Paper summary

DepState-JEPA extends V-JEPA 2.1 with separate RGB and depth encoders, metric positional encoding, and a shared latent state, DepState. Shared features are aggregated into the recurrent state, while modality-private features provide visual decoding memory. The same spatial readouts support masked-feature completion and action-conditioned future-feature prediction. Robot post-training adds state transitions, proprioceptive and object-motion supervision, and joint–TCP forward-kinematics consistency. The paper evaluates representation transfer, feature prediction, simulated control, and frozen sim-to-real visual-goal planning. The main contribution is a unified state interface, with the most informative empirical evidence coming from the geometry and future-state-supervision ablations on real manipulation.

Critical strengths
  1. The geometric state interface is coherent and has useful controlled evidence on real manipulation. The design connects metric RGB-D observations to a state used for action-conditioned prediction, while retaining separate visual decoding memory. Importantly, the depth-content/XYZ experiment retains two active backbones and the RGB-D teacher supervision.

  2. The real-robot results provide evidence about what makes the learned state useful for planning. The models are reported to be trained in simulation and deployed with frozen weights, and the full-task criteria include ordered execution and stable release. Removing future-state matching reduces drawer/cup success, despite retaining future-feature and physical supervision. In contrast, removing private decoding worsens drawer/cup feature-prediction, while real success remains. These results suggest that feature fidelity and control utility do not change in lockstep. The FK controls are also informative. These ablations already help explain several auxiliary objectives.

Critical weaknesses
  1. The real-robot RGB-D input provenance and registration require verification because they bear directly on the geometric construction. In the supplementary View 1 videos, RGB and colorized depth exhibit a substantial field-of-view and scale mismatch, affecting the platform, table boundary, and robot arm. Exporting native streams before registration is a plausible explanation, and the corresponding motion is broadly consistent. However, the submission does not establish how these displays relate to the actual registered inputs used for control. This is a consequential uncertainty: if registration occurs before model input, it should be straightforward to document; if the displayed pixel correspondences are used directly, the metric lifting and alignment interpretation would be compromised. I need an auditable raw-to-model-input example to assess this part of the method.

  2. The distinctive selective-sharing design is less well established than the benefits of geometry and state supervision. Preserving modality-specific decoding information is a reasonable motivation. The less clear point is whether restricting alignment to the shared subspace, together with the additional separation constraints, is responsible for the reported advantage over a simpler fused state. Ex.5 adds private-feature alignment without reducing the existing shared alignment, changing both alignment scope and total strength; Ex.4 changes both private information access and active decoder capacity (App. E.7.2). These comparisons do not isolate the central mechanism. Cross-method results also mix different representation adaptation, encoder-update policies, and supervision (App. E.3/E.5). Finally, the real-robot CEM objective is described as a terminal latent distance without specifying whether it uses DepState or decoded features, making the private pathway's role in actual candidate scoring unclear. A targeted, matched-strength alignment comparison and an explicit planning objective would substantially improve the interpretation.

Questions for authors
  1. RGB-D provenance and registration. Are the supplementary View 1 streams native outputs before registration? Which camera/view actually supplies the controller? Please provide a short continuous segment from one demonstration with raw RGB, metric depth, camera/frame IDs and acquisition timestamps, the relevant intrinsics/extrinsics and depth scale, and the registration/crop/resize procedure. Please also show the corresponding postprocessed RGB-D pair and updated intrinsics actually passed to the model.

  2. Real-robot planning objective. Please give the complete CEM cost: the prediction and goal representations, normalization, aggregation over tokens/layers/modalities, and whether private-conditioned decoding participates in candidate scoring. Which learned modules are invoked during each candidate rollout? This definition is necessary to interpret the shared-state and private-decoding results.

  3. Selective sharing. Does shared-only alignment still outperform full-space alignment when total alignment strength is held fixed? This is the specific comparison for which the current explanation and Ex.5 leave an ambiguity. A matched comparison, or a narrower interpretation of the existing result, would be more informative than an exhaustive sweep of loss removals. Please also clarify the strongest multimodal baseline's trainable modules and physical supervision during robot adaptation, so that the architectural comparison can be assessed.

8 comments
AuthorsOpenReview ↗

[Q1 | PART 1/4] CAMERA ROLES AND RGB-D PROVENANCE

[Q1 | PART 1/4] CAMERA ROLES AND RGB-D PROVENANCE

We thank the reviewer for raising questions about the real-robot camera setup and RGB-D acquisition pipeline. Below, we clarify the camera roles, distinguish the raw visualizations from the registered policy inputs, and provide the relevant calibration, data, and implementation details.


1. Camera Roles

As specified in Appendix B.1 (line 808), the policy uses only View 1, the fixed third-person Intel RealSense D435 camera, during both training and inference.

View 2 is an Intel RealSense D405 mounted on the Pika gripper from AgileX Robotics. It is recorded solely for supplementary visualization and provides a wrist-mounted view of local depth changes during manipulation.

The policy does not use View 2 as an input. All visual observations supplied to the policy come from the fixed third-person D435 (View 1).


2. Raw Visualization versus Actual Policy Inputs

We confirm that the RGB-D streams shown in Supplementary View 1 are the original, unregistered camera streams; they are not the registered RGB-D inputs supplied to the policy.

During robot execution, our recording script saves two distinct types of data:

For the supplementary demonstrations, we visualized the original .bag recordings rather than the registered RGB-D frames saved by the inference pipeline.

To render the raw depth stream for visualization, we played back and exported the .bag recordings using the official Intel RealSense Viewer with the Default preset. This preset displays the raw depth stream as a colorized visualization.

The following anoymous screenshot link shows the exact visualization configuration used:

RealSense Viewer — Default Preset Visualization : https://anonymous.4open.science/r/ICLR-9603-DepStateJEPA-Materials/4-RawCamera-RealsenseViewer/RawCamera-RealsenseViewer.png

Accordingly, the apparent spatial misalignment in Supplementary View 1 is attributable to its use of raw, unregistered streams. This visualization choice explains the apparent spatial misalignment and does not reflect the actual inputs used during policy inference.

AuthorsOpenReview ↗

[Q1 | PART 2/4] CAMERA CALIBRATION AND INTRINSICS

[Q1 | PART 2/4] CAMERA CALIBRATION AND INTRINSICS

3. Camera Calibration and the Appearance of View 2

The difference in appearance between View 1 and View 2 is consistent with their different imaging configurations.

Specifically, the Intel RealSense D405 mounted on the Pika gripper from AgileX Robotics uses the left imager of its stereo depth-sensing module to provide the color stream, rather than a separate, independently positioned RGB camera. Its depth stream is derived from the stereo imaging system. In contrast, the D435 uses separate color and stereo-depth imaging modules.

This hardware distinction is consistent with the raw color and depth visualizations of the D405 in View 2 appearing more spatially aligned than those of the D435 in View 1.

The RealSense Viewer screenshot (https://anonymous.4open.science/r/ICLR-9603-DepStateJEPA-Materials/4-RawCamera-RealsenseViewer/RawCamera-RealsenseViewer.png) also shows this distinction: the D405 highlighted by the red box appears as a single imaging module, whereas the D435 has separate color and depth imaging modules.

Camera Intrinsics

The table below reports the native camera intrinsics and the corresponding intrinsics after depth-to-color registration. All values correspond to the 640 × 480 stream profiles used in our experiments.

ViewCameraStream / Configurationfxfycxcy
View 1D435Native Color607.01196606.85699333.15720247.44922
View 1D435Native Depth386.72366386.72366317.56564245.77191
View 1D435Registered Color607.01196606.85699333.15720247.44922
View 1D435Registered Depth (aligned to Color)607.01196606.85699333.15720247.44922
View 2D405Native Color390.85461390.24884322.37259234.84363
View 2D405Native Depth388.47577388.47577321.48004235.45406
View 1D435Policy RGB (384 × 384)485.60956485.48560202.52576197.95938
View 1D435Policy Depth (384 × 384)485.60956485.48560202.52576197.95938

For View 1, the registered color and depth streams share the color-camera pixel coordinate system and therefore have identical intrinsics. The final two rows report the corresponding intrinsics after the policy's center-cropping and resizing operations. This matrix is stored as K in every policy-input NPZ file.

Note: View 2 is used exclusively for supplementary visualization and is not part of the policy-input pipeline. We therefore report only its native camera intrinsics.

AuthorsOpenReview ↗

[Q1 | PART 3/4] ADDITIONAL DATA AND IMPLEMENTATION FOR INDEPENDENT VERIFICATION

[Q1 | PART 3/4] ADDITIONAL DATA AND IMPLEMENTATION FOR INDEPENDENT VERIFICATION

4. Additional Raw Data and Implementation for Independent Verification

Regarding the request for original camera recordings, we would like to transparently clarify a data-retention issue.

The original recordings were made directly on the workstation used for policy inference, where available storage capacity was already significantly constrained. A single RGB-plus-depth .bag recording typically occupies approximately 5 GB per minute. After the original recording session, we deleted the corresponding .bag files to reclaim storage space. Therefore, unfortunately, the original .bag files corresponding to the submitted supplementary videos are no longer available.

New Recordings for Independent Verification

To address the reviewer's concern for verifiable evidence, we have performed new recordings and prepared two separate anonymous Hugging Face datasets:

  1. Raw Camera Recordings: Anonymous Hugging Face Dataset — Raw Camera Recordings.
  2. Registered Policy Inputs: Anonymous Hugging Face Dataset — Registered Policy Inputs.

The two datasets separately provide the original camera recordings and the corresponding registered RGB-D data used as policy inputs.

Camera calibration information: The .bag files preserve the camera stream information and calibration parameters, which can be inspected directly by opening them in the official Intel RealSense Viewer.

Publicly Available Acquisition and Preprocessing Code

We also provide an anonymous code repository containing the complete acquisition and preprocessing implementation:

Anonymous Repository — RealSense RGB-D Tools: (https://anonymous.4open.science/r/ICLR-9603-DepStateJEPA-Materials/3-RealSense-RGBD-Tools/README.md)

The repository includes:

ScriptDescription
record_realsense_rgbd.pyCamera recording script
realsense_bag_to_video.pyRaw .bag visualization and video conversion script
realsense_depstate_policy_input.pyPolicy-input acquisition and preprocessing script
Camera calibration codeCamera calibration and relevant implementation details

Together, the provided data and scripts document the acquisition and preprocessing pipeline, including:

  • Raw RGB and metric depth.
  • Camera/frame IDs and acquisition timestamps.
  • Relevant camera intrinsics and extrinsics.
  • Depth scale.
  • Registration, cropping, and resizing procedures.

For clarity, these are newly acquired recordings provided to facilitate independent verification, rather than recovered copies of the original deleted recordings.

Reviewer_NqLXOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper addresses the core gap between geometric scene understanding and action-driven state evolution in predictive visual representations for robot manipulation, and proposes DepState-JEPA, a dual-stream RGB-D joint-embedding predictive architecture. Built on the V-JEPA 2.1 backbone, the method anchors bimodal features into metric 3D space using measured depth and camera intrinsics, and aggregates a unified depth-grounded shared latent state (DepState) via a shared-private feature factorization strategy, while retaining modality-specific details for decoding. The framework supports both ray-conditioned spatial completion and action-conditioned temporal prediction, and introduces soft forward kinematics (FK) consistency regularization to improve the physical plausibility of states. Training follows a two-stage protocol: RGB-D representation adaptation and robot post-training. The paper validates the method across five task categories: spatial representation transfer, RGB-D feature completion, multi-step temporal forecasting, simulated control, and Sim2Real real-robot transfer. Systematic ablation studies quantify the contribution of each core component, and the overall performance outperforms existing mainstream JEPA baselines.

Critical strengths

Architectural innovation with paradigm-level reference value

Rather than simply adding a depth input to JEPA, this work proposes a novel multimodal representation principle: geometry grounding first, state unification second. It aligns two modalities in metric 3D space, then aggregates a unified DepState as the interface for both spatial querying and action prediction via shared-private factorization. This design bridges geometric perception and temporal prediction, and provides a clear architectural reference for the multimodal JEPA and robotic world model communities, representing a substantive architectural advance rather than an incremental improvement.

Comprehensive experimental pipeline with outstanding reproducibility

The evaluation covers the full pipeline from representation quality to downstream control, spanning perception, prediction, and manipulation tasks. Sixteen ablation studies precisely dissect the contributions of geometry grounding, feature sharing, temporal modeling, and physical constraints. The appendix provides extremely detailed implementation specifications, including full architecture parameters, loss weights, optimization protocols, data processing rules, and complete training/inference pseudocode. Its reproducibility is at a high level among comparable works, and core conclusions are supported by a thorough chain of evidence.

Critical weaknesses

Systematic fairness concerns in baseline comparisons

Several core baselines (3D-JEPA-AC, LeMuMo-JEPA-AC, etc.) are action-conditioned adaptations implemented by the authors, with no official reference implementations. Moreover, pretraining datasets, data scales, and compute budgets differ significantly across methods, with no strictly controlled comparison under unified pretraining data and training resources. The observed performance gains cannot be fully attributed to the DepState architectural innovation, and may be confounded by data scale and adaptation tuning effort, which directly undermines the credibility of the core conclusions.

Severely limited real-world validation leaves the method’s scope unclear

Real-robot experiments are restricted to two low-complexity structured tabletop tasks (drawer pulling/closing and cup pick-and-place), with a fixed scene, single object, and small motion displacement. Whether the core designs — depth grounding and fixed private memory — retain their advantages in unstructured environments, large-displacement interactions, or multi-object complex scenarios remains unvalidated. The practical generalizability and deployment boundary of the method are ambiguous, weakening the work’s real-world impact.

Questions for authors

The action-conditioned baselines marked with † are all adapted by the authors. Could you clarify whether the number of training iterations, compute budget, and model parameter count of these baselines are strictly aligned with DepState-JEPA during action-conditioned post-training? Have the authors considered retraining all baselines on the identical LingBot-Depth pretraining data with equal training resources, to eliminate the confounding effects of pretraining data differences and adaptation effort on performance comparison?

Ablation Ex.14 shows negligible performance difference between recursively updated private memory and fixed private memory under the current experimental scenarios. What is the core trade-off that leads the authors to choose fixed private memory as the default scheme? Given that real robotic manipulation often involves more complex visual changes such as large displacements, object rotations, and new objects entering the field of view — which may undermine the applicability of fixed private memory — do the authors agree that it is preferable to adopt the more generalizable recursive update scheme rather than clinging to the marginal advantage of fixed memory in simple scenarios? Is it possible that the performance ranking of the two schemes will reverse in more complex manipulation scenarios?

The depth encoder is initialized with V-JEPA RGB pretrained weights by replicating single-channel depth to three channels. What is the core rationale for this design? Have the authors run control experiments with random initialization or a dedicated depth-pretrained encoder, to verify the actual benefit of RGB pretrained weights for the depth branch?

4 comments
AuthorsOpenReview ↗

APPRECIATION FOR YOUR ENCOURAGING ASSESSMENT AND PATIENCE

We are carefully preparing a point-by-point response, including clarifications of the experimental protocols and additional analyses to address the concerns raised. We will share our detailed responses and relevant experimental results as promptly as possible.

Sincerely thank the reviewer for the thoughtful and encouraging assessment of our work, as well as for the constructive questions and suggestions and greatly appreciate your patience and understanding.

AuthorsOpenReview ↗

[W1&Q1] MATCHED STAGE-B ADAPTATION AND SYSTEMATIC COMPARISON

[W1&Q1] MATCHED STAGE-B ADAPTATION AND SYSTEMATIC COMPARISON

We thank the reviewer for highlighting a more systematic fairness comparision. We agree that the Stage-A pretraining remain an acknowledged difference. We therefore compare two representative RGB-D baselines, LeMuMo-JEPA-AC and Dual-Stream V-JEPA 2.1-AC under matched Stage-B.


1. Original-AC and Physical-Supervision-Matched Baseline Variants

As explicitly clarified in App.E.5 and the Tab.2 caption, the † marker denotes our action-conditioned adaptation of baselines, originally a representation learner rather than an action-conditioned model. Its added transition uses the same robot split, prediction horizon, transition architecture, and optimization budget as V-JEPA 2.1-AC, while retaining LeMuMo's pretrained RGB-D representation and native feature dimensions.

LeMuMo-JEPA-AC uses one shared encoder with modality-specific stems and fusion tokens; its Original variant freezes the encoder/fusion module during robot adaptation. Dual-Stream V-JEPA 2.1-AC uses two independent encoders (RGB and three-channel normalized depth), followed by token concatenation and projection; its Original variant freezes both encoders and the fusion projection. Both Original variants train only the transition and feature projections, without physical heads, direct physical supervision, or FK. The PSM variants unfreeze the representation modules and add the same four physical heads (joint, velocity, TCP, object), direct labels, and calibrated FK consistency as DepState. Dual-Stream lacks metric XYZ grounding, shared-private factorization, and the 64-query state.

2. Matched Stage-B Adaptation Protocol

Both PSM baselines use DepState's 20K Stage-B updates (15K robot + 5K RGB-D replay; every fourth update is replay), 384 resolution, effective batch 128 (micro-batch 16), AdamW, peak learning rates 2e-6/5e-5, 1K warm-up with cosine decay, rollout horizon 1 for the first 2K robot updates then 3, matched physical-loss weights/FK warm-up, and Stage-A-initialized EMA (0.99925). Replay uses each method's native objective. Original variants use 20K robot updates without replay or physical supervision.

Real-Robot Full-Task Success

MethodDrawer↑Cup↑
LeMuMo-JEPA-AC Original75.6 ± 6.974.4 ± 8.4
LeMuMo-JEPA-AC PSM77.8 ± 5.176.7 ± 3.3
DS-JEPA-AC Original73.3 ± 3.370.0 ± 3.3
DS-JEPA-AC PSM75.6 ± 3.873.3 ± 3.3
DepState-JEPA-AC81.1 ± 5.178.9 ± 3.8

All methods use the same tasks, initial conditions, visual subgoals, hardware interface, and CEM settings. DepState's mean success exceeds LeMuMo PSM by 3.3/2.2 points and Dual-Stream PSM by 5.5/5.6 points on Drawer/Cup.

3. Model Capacity and Compute

MethodTotal ParamsTrainable ParamsReplay-only ParamsTraining UpdatesGPU-hours
LeMuMo-JEPA-AC Original316.2M2.0M020k20
LeMuMo-JEPA-AC PSM324.8M317.2M7.6M20k29
DS-JEPA-AC Original620.3M12.0M020k26
DS-JEPA-AC PSM725.2M621.4M103.8M20k53
DepState-JEPA-AC662.2M659.1M3.15M20k38

Dual-Stream PSM is the more direct two-trunk comparison: both systems use two encoders and matched Stage-B updates/physical supervision. DepState has fewer total Stage-B parameters (662.2M vs. 725.2M) and GPU-hours (38 vs. 53), with 5.5/5.6 points higher mean success. Its trainable-online count is slightly higher, so capacity is not identical.

This compares dense feature prediction with geometry-grounded shared-state prediction, further supports the effectiveness of our core design: grounding multimodal representations in metric 3D space and unifying them into a shared state for action-conditioned prediction.

4. Method-Specific Stage-A Pre-training

Given the limited revision window and substantial computational cost, we retain each baseline's existing pretrained checkpoint and match the subsequent Stage-B adaptation. A common-data comparison on LingBot-Depth would require retraining each baseline with 35K Stage-A updates followed by 20K Stage-B updates for every seed, making it a prohibitively expensive experiment to complete within the current revision period.

We will complete this common-data comparison in a future version. The additional PSM experiments provide a controlled comparison of Stage-B adaptation under matched update schedules and physical supervision, while differences in Stage-A pretraining remain an acknowledged limitation.

5. Manuscript Revisions

  • Section 4.1 / Real-Robot Results: Specify the baselines and report Original/PSM results.
  • Appendix E.5: Clarify the adaptation and document PSM modules, supervision, Stage-B settings, parameter/compute accounting, and Stage-A checkpoints.

We thank the reviewer for prompting a more systematic comparison. We hope the matched Stage-B adaptation results and explicit disclosure of Stage-A differences address W1 and Q1.

AuthorsOpenReview ↗

[W2&Q2] ROLLOUT-CONSTANT PRIVATE MEMORY AND EXPANDED REAL-ROBOT EVALUATION

[W2 & Q2] ROLLOUT-CONSTANT PRIVATE MEMORY AND EXPANDED REAL-ROBOT EVALUATION

We thank the reviewer for helping clarify the rollout-level role of private memory and suggesting additional real-robot scenarios to demonstrate DepState-JEPA in more complex, unstructured settings. We clarify the terminology and report two additional multi-object tasks.


1. Terminology Clarification: Rollout-Constant Private Memory

Our earlier term “fixed private memory” could imply that features remain unchanged throughout real-robot execution. We will replace it with “rollout-constant private memory” throughout the revised manuscript.

Within each imagined rollout, P0mP_0^m remains fixed as DepState evolves under candidate actions and predicted proprioception. Our real-robot controller uses receding-horizon MPC: CEM plans eight steps (300 candidates, 15 refinements), executes the first action, obtains a new RGB-D observation, re-encodes DepState and private features, and replans.

“Rollout-constant” means unchanged only within one imagined rollout; after each executed action, receding-horizon MPC obtains a new RGB-D observation and initializes fresh private features for the next rollout.

2. Why We Use Rollout-Constant Private Memory

Ex.14 evaluates prediction-driven memory updating: it updates private memory from the predicted shared state, Ph+1m=Um(Phm,S^h+1)P_{h+1}^m = U_m(P_h^m,\hat S_{h+1}), without future visual observations. This differs from the observation-based refresh in real-robot MPC. Our rollout-constant design preserves the modality-specific information encoded from the latest actual observation throughout the imagined trajectory, while the next MPC cycle incorporates newly observed visual information.

Private-Memory Ablation (five-horizon predicted-proprioception forecasting; mean ± sample SD)

VariantDrawer L1/CosCup L1/CosAggregate SR↑
Ex.14: Prediction-updated memory0.380 ± 0.024 / 0.849 ± 0.0200.424 ± 0.026 / 0.812 ± 0.02383.1 ± 2.9
Ex.16: Rollout-constant memory0.358 ± 0.016 / 0.868 ± 0.0110.414 ± 0.022 / 0.822 ± 0.01983.4 ± 2.9

Ex.16 gives better mean visual-forecasting scores, while aggregate task success is similar. We use rollout-constant memory as a simple decoding reference, not to claim that prediction-driven updating is universally inferior; Ex.14 does not test the observation refresh already provided by MPC.

3. Additional Real-Robot Tasks

We add two tasks to evaluate more complex, unstructured multi-object manipulation:

  • Task 3 — Cluttered Storing. The robot pulls open the upper drawer of a tabletop organizer by its handle. It then grasps the block from among the objects scattered around the organizer, lifts it clear of the surrounding clutter, and places it inside the open drawer.

  • Task 4 — Cup Stacking. The robot picks up the loose blue cup, centers it over the green cup at the top of an existing stack, and lowers it into place. It then picks up the loose red cup and places it over the blue cup, leaving both cups nested upside down in a single stack.

Anonymous rollouts links of extending evaluation : https://anonymous.4open.science/r/ICLR-9603-DepStateJEPA-Materials/5-Unstructed_RealTask_Rollouts/README.md

Real-Robot Full-Task Success

MethodDrawer Pull & CloseCup Pick-and-PlaceCluttered StoringCup Stacking
LeMuMo-JEPA-AC Original75.6 ± 6.974.4 ± 8.458.9 ± 6.964.4 ± 5.1
LeMuMo-JEPA-AC PSM77.8 ± 5.176.7 ± 3.361.1 ± 6.966.7 ± 3.3
DS-JEPA-AC Original73.3 ± 3.370.0 ± 3.352.2 ± 5.157.8 ± 7.7
DS-JEPA-AC PSM75.6 ± 3.873.3 ± 3.357.8 ± 5.163.3 ± 5.8
DepState-JEPA81.1 ± 5.178.9 ± 3.872.2 ± 5.175.6 ± 3.8

DepState has the highest mean success on all four tasks, exceeding the strongest PSM baseline. The larger gains on the two new tasks highlight the value of geometry-grounded shared-state prediction in cluttered, interactive scenes.

4. Manuscript Revisions

  • Section 3.4 / Appendix E.6.3: Replace “fixed private memory” with “rollout-constant private memory” and clarify that new private features are encoded from the latest RGB-D observation at each MPC replanning step.
  • Appendix E.7.3: Describe Ex.14 as prediction-driven private-memory updating without future visual observations, and report its comparison with Ex.16 using the revised terminology.
  • Section 4.4: Add both tasks, all Original/PSM results, per-task trial counts, and anonymous rollout videos.

We particularly appreciate the reviewer's insightful, technically informed reading of our work. These suggestions helped clarify the memory mechanism and better highlight the contribution of geometry-grounded shared-state prediction. We hope the clarifications and added scenarios address W2/Q2 and would appreciate reconsideration. In light of the clarifications and new results, we would be grateful if the reviewer would consider a higher rating.

AuthorsOpenReview ↗

[Q3] DEPTH-ENCODER INITIALIZATION

[Q3] DEPTH-ENCODER INITIALIZATION

We thank the reviewer for the helpful suggestion to clarify the depth-encoder initialization and directly evaluate its effect.


1. Initialization Rationale and Training Schedule

We initialize the RGB and depth encoders from the same V-JEPA 2.1 checkpoint, but keep their parameters independent and adapt them separately. Metric depth is log-normalized and replicated across three channels to match the input interface of the V-JEPA 2.1 image/video stems. The unclipped metric depth and validity map are retained separately for geometric lifting.

This choice reuses the pretrained ViT structure and weights as an initialization; it does not tie the RGB and depth encoders or assume that their learned features remain semantically identical. During the first 2,000 Stage-A updates, the pretrained encoder parameters are frozen except for the depth image/video stems, while the newly introduced modules remain trainable. Both encoders are then unfrozen and independently adapted on RGB-D data for the remainder of Stage A.

The RGB checkpoint provides an initialization for the depth branch, which is subsequently adapted with independent weights on normalized metric-depth observations and explicit 3D grounding.

2. Controlled Depth-Initialization Ablation

Following the reviewer's suggestion, we add a direct comparison between the current RGB-checkpoint initialization and random initialization of the depth encoder. Only the depth-encoder initialization changes. Both variants follow the same first-2K stem-only warm-up and subsequent encoder-unfreezing schedule. The RGB encoder, architecture, depth preprocessing, RGB-D training data, objectives, optimizer, freeze/unfreeze schedule, and evaluation protocol remain the same. Both variants use 35K Stage-A updates and the same 20K PSM Stage-B schedule. Depth-feature completion is evaluated against the same fixed depth-reference encoder, prepared independently using training-scene depth data.

Depth-Encoder Initialization Ablation

Depth initializationDepth-feature completion L1↓Depth-feature completion Cos↑
V-JEPA 2.1 RGB checkpoint (current)0.254 ± 0.0040.947 ± 0.004
Random initialization0.275 ± 0.0080.932 ± 0.006
Relative improvement7.64% relative reduction+0.015 absolute gain

3. Manuscript Revisions

  • Appendix B.1–B.2: Clarify the depth normalization, three-channel input construction, independent encoder parameters, metric-depth preservation, and Stage-A freeze/unfreeze schedule.
  • Appendix E.7: Add the random-initialization control and report depth-feature completion results with seed-level variability.

We sincerely thank the reviewer for their patience and for the three constructive suggestions, which have helped us clarify the experimental comparisons, sharpen the presentation of the memory mechanism, and extend the real-robot evaluation. Together, these changes make the role and contribution of geometry-grounded shared-state prediction clearer.

We would be grateful if the reviewer would consider raising the work a higher rating in light of the clarifications and additional evidence. Please feel free to reach out with any further questions; we would be delighted to continue the discussion.

Reviewer_UQ4MOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces DepState-JEPA, an RGB–depth predictive architecture built around a geometrically grounded shared latent state. It combines shared–private feature projections, spatial readouts, action-conditioned transitions, physical supervision, and forward-kinematics regularization. Experiments evaluate representation transfer, feature prediction, and robot control.

Critical strengths

The framework provides a coherent connection between RGB–depth representation learning and action-conditioned prediction. Evaluation is broad and includes controlled component ablations, simulation benchmarks, and real-robot transfer.

Critical weaknesses

The necessity of several architectural components is not convincingly established. Ray queries show no clear advantage over pixel queries, removing private decoding barely affects aggregate control success, and FK regularization provides a relatively small aggregate gain. The main long-horizon forecasting comparison uses recorded proprioception at every transition. This evaluates conditioned visual forecasting, but does not directly establish autonomous rollout stability. Predicted-proprioception ablations are reported separately, without an equivalent cross-method comparison. Representation adaptation and physical supervision differ across complete systems. Consequently, the comparisons do not clearly isolate the benefit of the proposed representation from additional training and supervision.

Questions for authors

How does a simpler RGB–depth fusion model perform under the same representation updates and physical supervision? Can the authors compare all methods using predicted proprioception and report uncertainty for the smaller component gains?

4 comments
AuthorsOpenReview ↗

[W1–W3] ARCHITECTURAL CLARIFICATIONS AND COMPONENT ROLES

[W1–W3] ARCHITECTURAL CLARIFICATIONS AND COMPONENT ROLES


1. W1: Ray-Conditioned Spatial Readout

Ray queries provide an intrinsic-aware parameterization of spatial queries. We transform image coordinates into normalized camera-ray directions using the camera intrinsics, whereas pixel queries use normalized image coordinates.

With fixed intrinsics, the coordinates are deterministically related but differ geometrically: pixels identify image locations, whereas rays represent calibrated viewing directions. Recomputing rays using the corresponding camera intrinsics expresses each query in the calibrated viewing-direction coordinates of the current camera.

Ray queries provide an explicit camera-calibrated viewing-direction interface for reading local shared features, while preserving the feature-prediction performance observed with pixel queries under the evaluated camera setup.

2. W2: Private Features for Modality-Specific Visual Decoding

Only shared features HmH^m are aggregated into DepState and passed to the action-conditioned transition. Private features PmP^m serve as modality-specific decoder memory. Their initial values remain fixed during temporal rollout; they neither enter the recurrent state nor condition the transition. Decoded RGB/depth features are auxiliary predictions, while CEM candidate ranking uses the terminal shared DepState.

Private Decoding Ablation

ConfigurationDrawer L1/CosCup L1/CosDrawer SRCup SR
Ex.4: Private residual disabled0.400 ± 0.021 / 0.842 ± 0.0170.456 ± 0.027 / 0.786 ± 0.02580.0 ± 6.777.8 ± 5.1
Ex.16: Full model0.358 ± 0.016 / 0.868 ± 0.0110.414 ± 0.022 / 0.822 ± 0.01981.1 ± 5.178.9 ± 3.8

Forecasting uses predicted proprioception after initialization. Removing private decoding worsens Drawer/Cup L1 by 0.042 and lowers Cos by 0.026/0.036, while real-robot full-task success remains similar. Ex.4 disables the private cross-attention residual but retains the query path, private projections, and factorization losses; it therefore ablates the private-conditioned decoding pathway, including private observation access and active private-attention capacity.

The private branch preserves modality-specific detail for visual reconstruction and prediction; recurrent transition and CEM action evaluation operate through the shared DepState.

3. W3: Role of Forward-Kinematics Consistency

The FK term couples independently supervised joint-configuration and TCP-pose predictions through calibrated robot kinematics. This is relevant to robot execution because small TCP-pose changes can sometimes lead to disproportionately large changes in IK-derived joint angles. Coupling the two predictions through calibrated robot kinematics helps prevent kinematically inconsistent joint/TCP predictions and supports physically consistent robot-state estimation. During training, FK gradients reach both physical readout heads and the latent transition, constraining the state recursively used at subsequent steps.

It provides a kinematic constraint on recursive physical-state prediction, rather than an objective that directly maximizes task success.

The relevant ablation uses measured proprioception only at initialization; predicted joint configurations and velocities are then recursively fed into subsequent transitions. Each Drawer/Cup entry averages the five forecasting horizons (1, 8, 16, 24, and 32 steps) over the same evaluation windows. Full FK reduces forecasting L1 versus no FK by 6.3% on Drawer and 7.4% on Cup; versus heads-only FK, it provides a further 3.5% and 4.2% reduction.

FK Ablation under Predicted-Proprioception Forecasting

ConfigurationDrawer L1/CosCup L1/Cos
Ex.12: No FK0.382 ± 0.024/ 0.849 ± 0.0190.447 ± 0.029/0.793 ± 0.026
Ex.13: FK to heads only0.371 ± 0.020/0.857 ± 0.0160.432 ± 0.027/0.806 ± 0.024
Ex.16: FK to heads and state0.358 ± 0.016/0.868 ± 0.0110.414 ± 0.022/0.822 ± 0.019

Under recursively predicted proprioception, full FK improves five-horizon visual forecasting over both no-FK and heads-only variants, with an additional benefit when consistency gradients reach the latent state.

4. Manuscript Revisions

  • Sections 3.2–3.4: Specify the camera-calibrated role of ray queries, distinguish the shared transition and private decoding pathways, and clarify FK coupling and its gradient propagation to the latent state.
  • Appendix E.7.2/E.7.4: Clarify the Ex.4 private-residual intervention and connect its Drawer/Cup forecasting results (tab17) with real-robot full-task results (tab18).
  • Appendix E.7.3/E.7.4: Detail the Ex.12/13/16 FK gradient pathways and clarify the five-horizon predicted-proprioception forecasting protocol in tab 17.

We thank the reviewer for considering these clarifications. We hope they address W1–W3 and would appreciate reconsideration of the assessment.

AuthorsOpenReview ↗

[Q1&W5] MATCHED RGB-D FUSION AND PHYSICAL SUPERVISION

[Q1&W5] MATCHED RGB-D FUSION AND PHYSICAL SUPERVISION

We thank the reviewer for highlighting the disentanglement the contribution of DepState's representation architecture from RGB-D fusion, adaptation updates, and physical supervision. We address this with a simple Dual-Stream(DS) baseline and Physical-Supervision-Matched (PSM) variants of it and LeMuMo-JEPA-AC.


1. Baseline Architectures and Original Adaptation

Dual-Stream-VJEPA2.1-AC. This baseline uses two independent V-JEPA 2.1 ViT-L/16 encoders initialized from the same RGB checkpoint: one receives RGB, the other normalized depth replicated across three channels. Corresponding tokens are concatenated and projected into a dense fused feature grid, which a frame-causal action-conditioned predictor forecasts using stop-gradient future RGB-D features as targets. It has no metric XYZ encoding, shared-private factorization, or learned 64-query state. Unlike LeMuMo, it uses simple token-wise fusion rather than native fusion tokens.

For DS-VJEPA2.1-AC (Original), both encoders and the fusion projection are frozen in Stage B; only the transition and feature projections are trained, without physical heads, direct physical supervision, or FK.

For LeMuMo-JEPA-AC (Original), it retains its native RGB-D fusion-token architecture and representation objective during Stage A. In Stage B, its encoder/fusion module is frozen; only the action-conditioned transition and feature projections are trained against future fusion-feature targets. It has no physical readout heads, direct physical losses, or FK regularization; actions and recorded proprioception are conditioning inputs.

2. Physical-Supervision-Matched Adaptation

For DS-VJEPA2.1-AC (PSM), we unfreeze both encoders and the fusion projection in Stage B, add the same four readout heads (joint configuration, joint velocity, TCP pose, object motion), and apply DepState's direct physical losses and calibrated joint-TCP FK consistency.

For LeMuMo-JEPA-AC (PSM), we retain its native fusion architecture but make its encoder/fusion module trainable in Stage B and add the same heads, physical losses, and FK constraint.

Both PSM variants use an EMA target initialized from Stage A (momentum 0.99925), 20K Stage-B updates, 25% RGB-D replay, and the same resolution, batch size, rollout curriculum, physical-loss weights, and FK warm-up as DepState. Replay retains each method's native representation objective. All methods use three seeds; Stage-A initialization and pretraining remain method-specific.

Adaptation Comparison

MethodStage-B representationFusion/statePhysical heads/losses
DS-VJEPA2.1-AC (Original)FrozenConcat + projection; dense gridNone
DS-VJEPA2.1-AC (PSM)TrainableConcat + projection; dense grid4 heads; direct losses + FK
LeMuMo-JEPA-AC (Original)FrozenNative fusion tokensNone
LeMuMo-JEPA-AC (PSM)TrainableNative fusion tokens4 heads; direct losses + FK
DepState-JEPA-AC (Ours)TrainableMetric XYZ; shared-private; 64-query state4 heads; direct losses + FK

3. Real-Robot Results

All methods use the same tasks, initial conditions, visual subgoals, hardware interface, and CEM budget; each task has 30 trials per seed. Entries are mean ± sample SD across three seeds.

MethodDrawer Full SRCup Full SR
DS-VJEPA2.1-AC (Original)73.3 ± 3.370.0 ± 3.3
DS-VJEPA2.1-AC (PSM)75.6 ± 3.873.3 ± 3.3
LeMuMo-JEPA-AC (Original)75.6 ± 6.974.4 ± 8.4
LeMuMo-JEPA-AC (PSM)77.8 ± 5.176.7 ± 3.3
DepState-JEPA-AC (Ours)81.1 ± 5.178.9 ± 3.8

Under matched Stage-B updates and physical supervision, DepState exceeds PSM LeMuMo by 3.3/2.2 pp on two tasks, and PSM dual-stream V-JEPA 2.1 by 5.5 pp on Drawer and 5.6 pp on Cup.

4. Manuscript Revisions

We will make the following targeted revisions:

  • Section 4.1: Describe DS-VJEPA2.1-AC and distinguish its dense token fusion from LeMuMo's fusion tokens and DepState's geometry-grounded state.
  • Appendix E.5 and the real-robot results table: Specify Original/PSM trainable modules, target branches, physical heads/losses, FK, replay and Stage-B budgets; report the baseline results and matched conditions.

We thank the reviewer for highlighting a controlled comparison of RGB-D fusion and adaptation. We hope these matched baselines address Q1 and W5 and would appreciate reconsideration of the assessment.

AuthorsOpenReview ↗

[Q2&W4] PREDICTED-PROPRIOCEPTION FORECASTING AND UNCERTAINTY

Q2 & W4: PREDICTED-PROPRIOCEPTION FORECASTING AND UNCERTAINTY

We thank the reviewer for highlighting this distinction. We extend the temporal comparison to recursive predicted-proprioception rollouts and report seed-level variability, complementing the existing recorded-proprioception evaluation.


1. Cross-Method Evaluation Protocol

In Table 2, each method receives recorded actions and proprioception at every transition, evaluating conditioned visual forecasting. In the complementary protocol, measured proprioception is used only at initialization; each model's predicted joint configuration and velocity then recursively condition subsequent transitions.

Each baseline receives the same four physical readout heads (joint configuration, joint velocity, TCP pose, object motion), direct physical supervision, and calibrated FK consistency. Heads operate on pooled features from each method's native representation. For more direct comparison, we show respresentive baselines PSM LeMuMo and DS-VJEPA2.1 following the configurations in our *Q1 / W5 response*. Stage-A initialization and representation objectives remain method-specific; the matched conditions concern Stage-B physical supervision and update budgets.

All methods start from the same RGB-D frame and measured proprioception, receive identical recorded actions and transition intervals, and then recursively use their own predicted joint configuration and velocity. Future proprioception and RGB-D are evaluation targets only. We score horizons 1, 8, 16, 24, and 32 on identical episodes and starting frames, using the same common RGB reference-feature space and endpoint-scoring procedure.

2. Predicted-Proprioception Forecasting

All baseline rows are PSM. Values are mean ± sample SD across three training seeds.

Method1-step L1 / Cos8-step L1 / Cos16-step L1 / Cos24-step L1 / Cos32-step L1 / Cos
DS-VJEPA2.1-AC0.148±0.006 / 0.970±0.0040.335±0.016 / 0.875±0.0150.469±0.024 / 0.792±0.0230.579±0.031 / 0.721±0.0280.674±0.037 / 0.652±0.035
LeMuMo-JEPA-AC0.151±0.005 / 0.970±0.0030.340±0.016 / 0.874±0.0160.476±0.021 / 0.797±0.0200.571±0.027 / 0.731±0.0240.650±0.031 / 0.675±0.032
DepState-JEPA-AC0.145±0.003 / 0.976±0.0010.319±0.014 / 0.893±0.0090.424±0.020 / 0.834±0.0170.494±0.027 / 0.781±0.0230.548±0.033 / 0.741±0.025

With identical actions and recursively predicted proprioception, DepState achieves the best L1 / Cos at every horizon. At 32 steps, its L1 is 15.7% lower than LeMuMo-JEPA-AC (PSM) (0.548 vs. 0.650), while its Cos is 0.066 higher than LeMuMo-JEPA-AC (PSM) (0.741 vs. 0.675).

3. Uncertainty of Component Effects

Appendix E.7 reports mean ± sample SD across three seeds. We additionally report paired seed-level differences, computed from unrounded results, between Ex.16 and:

  • Ex.7: Pixel vs. ray queries.
  • Ex.4: Private-conditioned decoding.
  • Ex.12: FK vs. no FK.
  • Ex.13: Full FK gradients vs. heads-only FK.
ComparisonL1 DifferenceCos Difference
Ex.16 vs. Ex.4-0.0503 ± 0.0055+0.0283 ± 0.0055
Ex.16 vs. Ex.12-0.0193 ± 0.0055+0.0157 ± 0.0050
Ex.16 vs. Ex.13-0.0107 ± 0.0031+0.0087 ± 0.0031
Ex.16 vs. Ex.70.0000 ± 0.0000-0.0003 ± 0.0006

The paired seed-level results show consistent visual-prediction improvements from Private Decoding and FK regularization, including the additional benefit of propagating FK gradients to the latent state. In contrast, Ray and Pixel queries exhibit negligible differences.

4. Manuscript Revisions

We will make the following targeted revisions:

  • Section 4.3 / Table 2: Identify the existing results as recorded-proprioception conditioned forecasting.
  • Appendix E.5: Add the cross-method recursive protocol, physical heads / losses and Stage-B settings, multi-horizon results, and clarify that future measurements are targets only and Stage-A pretraining remains method-specific.
  • Appendix E.7.4: Report paired seed-level differences and their variability for the component ablations.

We hope these clarifications and revisions adequately address the concern and would greatly appreciate reconsideration of the overall assessment. Please feel free to reach out if you have any further questions. We would be delighted to continue the discussion.

AuthorsOpenReview ↗

[FOR REVIEWER UQ4M] FOLLOW-UP ON OUR REBUTTAL AND ADDITIONAL EVIDENCE

[FOR REVIEWER UQ4M] FOLLOW-UP ON OUR REBUTTAL AND ADDITIONAL EVIDENCE

Dear Reviewer UQ4M,

We hope you are doing well. We would like to briefly follow up on our rebuttal and thank you for your patience during the discussion. Your three suggestions have helped us further clarify the architectural motivations and strengthen the experimental evidence for DepState-JEPA.

1. Architectural Clarifications (W1–W3).

We clarified the distinct purposes of the three components. Ray queries use camera intrinsics to parameterize calibrated viewing directions, providing a geometry-aware interface for spatial feature readout rather than claiming an accuracy advantage over pixel queries. The shared features construct DepState and drive action-conditioned transitions, whereas private features preserve modality-specific information exclusively for visual decoding. FK couples independently predicted joint configurations and TCP poses through calibrated robot kinematics, enforcing physical consistency during recursive state prediction rather than directly optimizing task success.

2. Baseline Fairness (Q1 & W5).

We introduced a dual-stream RGB-D baseline and physical-supervision-matched variants of both Dual-Stream and LeMuMo-JEPA-AC. By aligning Stage-B adaptation and physical supervision, these comparisons help isolate the contribution of geometry-grounded state construction from conventional RGB-D fusion and additional physical supervision.

3. Recursive Forecasting and Uncertainty (Q2 & W4).

We extended the cross-method evaluation from recorded-proprioception-conditioned forecasting to recursive rollouts using each model's predicted proprioception. We also added seed-level uncertainty analysis for component ablations. These results provide further evidence of DepState's long-horizon predictive consistency under autonomous proprioceptive feedback.

Overall, these additional analyses clarify the central design principle of DepState-JEPA: measured depth and camera intrinsics provide metric 3D grounding for RGB-D features; shared features are then aggregated into a unified state for spatial reasoning and action-conditioned prediction, while private features retain modality-specific visual details. This geometry-grounded state construction distinguishes our approach from conventional feature-level RGB-D fusion and provides a common representation for perception and manipulation.

We appreciate your careful reading and patience throughout this discussion. Your suggestions have helped us make the architectural distinctions clearer and better demonstrate where the contributions of our work lie. If these clarifications and additional results address your concerns, we would sincerely appreciate your reconsideration of the overall assessment.

Thank you again for your time and constructive engagement.

Comments on the paper

2 comments
AuthorsOpenReview ↗

SINCERELY THANK THE REVIEWER YXmz FOR THE CAREFUL, CLEAR, AND RESPONSIBLE ASSESSMENT

We sincerely thank the reviewer for the careful, clear, and responsible assessment. Your constructive comments have helped us identify important issues and improve the clarity and quality of our work. We greatly appreciate your reconsideration and revised assessment.

We also appreciate your suggestion regarding the supplementary visualizations. We will update spatially aligned RGB-D visualizations into the final version and provide updated supplementary rollouts with aligned RGB and depth streams.

Thank you again for your thoughtful and constructive engagement.

AuthorsOpenReview ↗

SUMMARY OF MANUSCRIPT REVISIONS AND CLARIFICATIONS

SUMMARY OF MANUSCRIPT REVISIONS AND CLARIFICATIONS

We have further revised the manuscript in response to the reviewers' comments. The main updates are summarized below:

1. Methodological Clarifications (Introduction; Sections 3.2–3.5; Appendices B.1–B.2 and E.6.3). We clarified the geometry-grounded construction of DepState, the distinct roles of shared-state aggregation and modality-private visual decoding, and the use of camera-calibrated ray queries for spatial readout without target depth. We also clarified the rollout-constant private-memory mechanism under receding-horizon MPC, the registered RGB-D inputs and calibration pipeline, and the terminal latent-distance objective used in CEM planning.

2. Controlled Comparisons and Additional Analyses (Sections 4.3–4.5; Appendices E.5–E.7). We added a coefficient-sum-matched selective-sharing comparison, Stage-B physical-supervision-matched RGB-D baselines, recursive forecasting with predicted proprioception, a depth-encoder initialization ablation, and seed-level uncertainty analyses. We also explicitly distinguish matched Stage-B adaptation from method-specific Stage-A pretraining.

3. Expanded Real-Robot Validation (Section 4.4; Appendix E.6.3). We extended the real-robot evaluation to Cluttered Storing and Cup Stacking, reporting both Original and PSM baselines alongside DepState-JEPA.

Taken together, these revisions help us further clarify the contribution of DepState-JEPA: RGB provides semantic appearance (what things are), while measured depth and camera intrinsics establish metric 3D space geometry (where they are). This motivates our design principle: Geometry grounding first, State unification second.

Guided by this principle, measured depth and camera intrinsics ground both streams in a common metric 3D coordinate frame. Selective shared-private factorization aggregates cross-modal shared features into DepState while preserving modality-specific details for decoding. DepState supports spatial readout through camera-parameterized viewing-ray queries and action-conditioned prediction of future visual and physical states, connecting spatial completion to visual-goal-conditioned MPC and robotic manipulation.

All corresponding revisions in the revised manuscript are highlighted in a soft lavender-blue font for clarity.

We welcome any further questions or discussion from reviewers who have not yet had time or opportunity to respond.

If the clarifications and additional evidence address your concerns, we would appreciate your reassessment of the paper's rating in light of the revised manuscript.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.