PaperlayerView exact revision on arXiv

Source-linked AI summary

PRISM: Learning Realistic Depth via Physics-Grounded Noise Disentanglement with Semantic-Geometric Collaboration

Xiujian Liang, Jiacheng Liu, Mingyang Sun, Qichen He, Anda Cheng, Cewu Lu, Jianhua Sun

arXiv:2610.29432v1cs.RO

TL;DR

Depth sensing’s heterogeneous noise complicates sim-to-real transfer, while existing models often conflate distinct physical error sources. PRISM uses 3D VFM semantic priors to disentangle and synthesize those errors, achieving 92.1% average real-world robotic success and state-of-the-art noisy-depth fidelity.

  • Problem

    Depth sensing exhibits heterogeneous noise, and existing models often treat distinct physical error sources as one monolithic process.

    Source: §1 · pp. 1–2

  • Method

    PRISM uses 3D VFM semantic-physics priors to disentangle Sensing Invalidation from Measurement Inaccuracy and guide noise synthesis.

    Source: §1 · p. 2

  • Results

    92.1% average success rate in the real world, a 53.2% absolute improvement over raw simulation, alongside state-of-the-art fidelity in noisy depth synthesis.

    Source: §1 · p. 2

  • Takeaways & Limitations

    PRISM demonstrates that physics-grounded disentanglement with semantic reasoning can support real-world robotic success and high-fidelity noisy depth synthesis.

    Source: §1 · p. 2, §6 · p. 9

    PRISM is trained for a fixed sensor family, lacks explicit temporal consistency for dynamic scenes, and does not model interaction-induced extrinsic noise or extreme optical environments.

    Source: §7 · p. 9

Abstract

from arXiv · show

Real-world physical sensing exhibits complex, heterogeneous noise patterns that deviate significantly from idealized simulation, posing a fundamental bottleneck for sim-to-real transfer. Existing sensor modelings typically treat depth noise as a monolithic black-box process, overlooking the distinct physical mechanisms that govern different error modalities. In this work, we introduce a physics-grounded paradigm that disentangles monolithic noise into two complementary modalities: \textbf{sensing invalidation} and \textbf{measurement inaccuracy}, enabling a tailored treatment of noise sources based on their physical origins. Building on this insight, we propose \textbf{PRISM}, a tripartite framework that distills 3D Visual Foundation Model features as rich spatial-semantic priors for physics-based reasoning. To address the inherent sparsity and class imbalance of invalidation regions, we develop Hierarchical Positive-Prioritized Supervision, integrating multi-scale positive-weighted objectives with a positive-preserving dynamic hard mining strategy to enforce precise artifact delineation. Extensive benchmarks demonstrate that PRISM achieves state-of-the-art fidelity in noisy depth synthesis. Furthermore, downstream robotic experiments show that PRISM facilitates a \textbf{92.1\%} average success rate in the real world, marking a significant improvement over monolithic baselines. Our project page is available at: \resizebox{\linewidth}{!}{\href{https://shawnleung-sii.github.io/prism-icml2026/}{https://shawnleung-sii.github.io/prism-icml2026/}.}

1. Introduction

PRISM treats depth noise as distinct physical error modalities rather than one stochastic process. Its semantic-geometric framework targets each modality separately, with positive-prioritized supervision for sparse invalidation regions.

1. Introduction

  • Prior monolithic models can hallucinate noise in clean regions or miss structural voids because they conflate distinct physical mechanisms.This limits reproduction of spatially correlated sensing artifacts.Source: §1 · p. 2
  • Real-world depth noise includes Sensing Invalidation from specular reflection and Measurement Inaccuracy from refraction or subsurface scattering.The paper treats these as complementary, physically motivated modalities requiring specialized modeling.Source: §1 · p. 2
  • PRISM’s Reason → Disentangle → Generate pipeline uses frozen 3D VFM features as semantic-physics priors, predicts pixel-wise invalidation, then synthesizes scene-consistent residual maps.The Semantic-Physics Reasoner supplies priors to the Bimodal Noise Disentangler, which conditions the Noise Residual Generator.Source: §1 · p. 2
  • Hierarchical Positive-Prioritized Supervision addresses sparse invalidation regions with multi-scale positive-weighted objectives and positive-preserving dynamic hard mining.Invalidation regions are often under 10% of pixels, making standard losses vulnerable to class imbalance.Source: §1 · p. 2
  • 92.1% average success rate in the real world, with a 53.2% absolute improvement over raw simulation, accompanies PRISM’s state-of-the-art fidelity in noisy depth synthesis.The robotic result is reported alongside extensive benchmark results for noisy depth synthesis.Source: §1 · p. 2

2. Related Work

Related work frames depth sensing as a key perceptual gap in sim-to-real transfer, reviews analytical and data-driven noise modeling, and motivates VFM features as semantic-physical priors.

2.1. The Perceptual Gap in Sim2Real Transfer

  • Offline, data-centric sensor modeling embeds real-world noise in training data, allowing standard policies to learn artifact robustness without online computational overhead.The cited policy-centric alternatives rely on deployment-time adaptation modules and may face latency and generalization bottlenecks.Source: §2.1 · p. 2

2.2. Paradigms of Depth Noise Modeling

  • Analytical modeling ranges from rigorous but costly ray-tracing to lightweight heuristics that lack material-dependent physical fidelity.Data-driven methods can look realistic but often conflate specular signal loss and transparent-material bias into monolithic noise.Source: §2.2 · p. 3

2.3. Visual Foundation Models as Semantic Priors

  • PRISM uses VFM-derived semantic-physical correspondences to guide physics-grounded noise disentanglement and material-aware artifact synthesis.The authors describe this as the first use of VFM-derived priors for this purpose.Source: §2.3 · p. 3

3. Methodology

PRISM models real-world depth corruption as sensing invalidation and measurement inaccuracy, then uses a semantic-to-geometric pipeline to reason about and synthesize these modalities. Its supervision strategy prioritizes sparse invalidation regions while training the disentangler and residual generator in sequence.

3. Methodology

  • PRISM formulates depth corruption as sensing invalidation plus measurement inaccuracy, learning an invalidation mask and a residual conditioned on the input image, simulated depth, and mask.The mask marks invalid pixels; the residual represents measurement error, and the two modalities are learned through physics-grounded disentanglement.Source: §3 · p. 3

3.1. Semantic-Physics Reasoner

  • SPR distills frozen 3D-VFM patch features into global semantic context and projects that context into multi-scale prompts for downstream geometric processing.The prompts are scale-specific and map to decoder channel dimensions.Source: §3.1 · p. 3

3.2. Bimodal Noise Disentangler

  • BND predicts a full-resolution invalidation probability map, preserving fine geometric cues and injecting SPR priors into its coarse-to-fine decoder.Its predicted map also confines residual generation to valid sensor regions.Source: §3.2 · p. 4

3.3. Noise Residual Generator

  • NRG uses semantic-conditioned, depth-controlled latent denoising to generate measurement residuals, with its training loss restricted to valid sensor regions.A frozen VAE and denoising U-Net work with a trainable ControlNet adapter; the decoded residual is added to simulated depth.Source: §3.3 · p. 4

3.4. Hierarchical Positive-Prioritized Supervision

  • Hierarchical Positive-Prioritized Supervision counters sparse invalidation labels with multi-scale positive weighting, positive-preserving hard-negative mining, and sequential BND/NRG objectives.The mining ratio decays from 0.25 to 0.1; Stage I adds Dice loss for shape compactness, while Stage II trains NRG with frozen BND and a valid-region mask.Source: §3.4 · p. 5

4. Application

The application uses PRISM as an offline enhancer for simulated depth, then trains a depth-aware policy on the enhanced demonstrations. The policy is deployed on physical robots without online domain adaptation.

Data Generation Pipeline

  • PRISM processes clean simulated depth offline to create real-world-like corrupted observations before policy training.The pipeline collects expert demonstrations in simulation, enhances their depth, and trains on the resulting dataset.Source: §4 · pp. 5–6

Qualitative Evaluation

  • Across specular and transparent scenes, PRISM separates invalidation masks from continuous residuals and localizes artifacts to physically unstable regions.The figure presents this as a qualitative comparison of disentangled generation and baseline outputs.Source: Figure 4 · p. 6

Policy Learning & Zero-Shot Transfer

  • A dual-stream RGB-and-depth policy is trained on PRISM-enhanced data and deployed zero-shot, without online adaptation.The depth branch is a dedicated encoder, and the policy is trained entirely on the enhanced dataset.Source: §4 · p. 6

5. Experiment

Experiments evaluate PRISM’s depth synthesis fidelity, cross-camera generalization, and real-world robotic performance. Results and ablations support its physics-grounded, semantically informed noise modeling and positive-prioritized supervision.

5.1. Experimental Settings

  • The evaluation uses ByteCameraDepth for training, NYU-Depth-v2 for zero-shot testing, and two robot platforms for six manipulation tasks.The platforms are Franka Research 3 for specular and deformable tasks and Realman RM75 for long-horizon articulation tasks.Source: §5.1 · p. 6, Figure 5 · p. 7, §5.3 · p. 8

5.2. Depth Fidelity Evaluation

  • PRISM achieves state-of-the-art ByteCameraDepth fidelity, with high invalidation IoU/F1 and the lowest valid-region MAE/RMSE among compared methods.Qualitatively, it targets artifacts on specular and transparent surfaces while preserving valid geometric structures.Source: §5.2 · p. 7, Table 1 · p. 7
  • Zero-shot tests on NYU-Depth-v2 show robust synthesis under a Kinect camera domain shift, with minimal performance degradation relative to overfitting-prone baselines.The authors attribute generalization to generic VFM priors and material-correlated visual patterns rather than dataset-specific textures.Source: §5.2 · p. 7, Table 2 · p. 7

5.3. Downstream Application Evaluation

  • 92.5% average success rate across six real-world tasks: PRISM significantly outperforms analytical and generative simulation baselines using the RISE policy.The tasks span specular, deformable, articulated, and long-horizon manipulation across two robotic platforms.Source: §5.3 · p. 8
  • PRISM improves Sim2Real success by about 3% over the strongest generative baseline across DP3, PPT, and RISE policies.The cross-policy comparison reports average success across all six tasks.Source: §5.3 · p. 8, Table 4 · p. 8

5.4. Ablation Studies

  • Removing SPR reduces performance on non-Lambertian surfaces, while removing BND produces the largest overall-MAE degradation and severe hole hallucinations.The NRG-only baseline has the lowest robotic success rate.Source: Figure 6 · p. 8, §5.4 · p. 9
  • Geometry-aware 3D VFMs outperform generic ViTs and CNNs, while MoGe2 achieves the highest Robot SR at greater parameter overhead.The authors suggest geometric foundation models encode transferable semantics-physics correlations.Source: §5.4 · p. 9, Figure 7 · p. 9
  • Multi-scale positive weighting addresses sparse invalidation masks, and dynamic PP-DHM improves Boundary-IoU by 8.8% (12.0% relative) over static mining.Dice Loss improves predicted-hole compactness; the dynamic decay strategy stabilizes training and sharpens artifact boundaries.Source: §5.4 · p. 9, Figure 8 · p. 9

6. Conclusion

The conclusion presents PRISM as a framework for disentangling depth-sensor noise through physics-grounded generation and VFM-driven reasoning. It argues that embedding physics and semantics into noise modeling can bridge the geometric perceptual gap for embodied AI.

6. Conclusion

  • PRISM reframes depth sensor noise through disentanglement to expose its physical causality.The framework combines VFM-driven reasoning with physics-grounded generation.Source: §6 · p. 9
  • The conclusion contrasts physics-semantic noise modeling with black-box randomization as an approach to the geometric perceptual gap.It presents this direction as a path toward trustworthy embodied AI.Source: §6 · p. 9
  • The proposed scope is embodied AI, where the paper links noise modeling to bridging the geometric perceptual gap.Source: §6 · p. 9

7. Limitations

PRISM’s limitations include restricted transfer across sensor families, no explicit temporal consistency for dynamic scenes, and a focus on intrinsic surface-correlated noise. The appendix documents its architecture, training methods, and inference pipeline.

7. Limitations

  • PRISM’s sensor-specific training degrades under out-of-distribution sensing principles; its per-frame generation also leaves temporal flicker in dynamic scenes unaddressed.The formulation additionally excludes interaction-induced extrinsic noise and domain-shifted extreme optical environments.Source: §7 · p. 9

Appendix Overview

  • The appendix provides architecture specifications, algorithmic pseudocode, experimental settings, dataset descriptions, extended evaluations, and qualitative visualizations.It also includes limitations and broader societal impact discussion.Source: Appendix Overview · p. 12

A.2.1. SEMANTIC-PHYSICS REASONER (SPR)

  • The SPR uses global average pooling and an MLP to turn frozen VFM features into semantic embeddings, then projects them into scale-specific channel dimensions.The VFM backbone is frozen during training.Source: §A.2.1 · pp. 12–13, Table 5 · p. 13

A.2.2. BIMODAL NOISE DISENTANGLER (BND)

  • The BND predicts pixel-wise sensing-invalidation probabilities, retaining full-resolution encoder features before downsampling for boundary delineation.Its decoder uses SPR priors to modulate skip-connected features channel-wise.Source: §A.2.2 · p. 13

A.2.3. NOISE RESIDUAL GENERATOR (NRG)

  • The NRG uses a trainable ControlNet adapter with a frozen Stable Diffusion backbone to synthesize depth-noise residuals.The appendix also specifies DDPM/DDIM settings and log-scale depth normalization.Source: §A.2.3 · p. 14

A.3. Algorithmic Description

  • PP-DHM preserves every positive pixel and mines only hard negatives, with dynamic decay shifting from broader coverage toward boundary-focused learning.This counters standard OHEM’s loss of rare positive examples and resulting all-negative predictions.Source: §A.3 · p. 14
  • PRISM’s inference pipeline extracts VFM semantic features, predicts an invalidation mask, generates a residual with conditional diffusion, and composes the synthesized depth.The pipeline applies the residual only outside the predicted invalidation mask.Source: §A.3 · p. 15

A.4. Training Hyperparameters

The appendix specifies PRISM’s training configuration and H-PPS objective, including positive-weighted supervision and positive-preserving hard mining. NRG training proceeds in two phases, first aligning the adapter and then refining the generator.

A.4. Training Hyperparameters

  • PRISM is trained on 8 H200 GPUs; BND uses 100 epochs, while NRG trains ControlNet for 50 epochs before 50 epochs of joint fine-tuning.The joint phase uses a reduced learning rate of 1 × 10−5.Source: §A.4 · p. 15

B.1. Implementation Details

  • H-PPS combines positive-weighted BCE, multi-scale mask supervision, and dynamic hard-negative mining while preserving all positive pixels.Its positive-class weight is 3.0, and the loss explicitly penalizes missed invalidations to support high recall.Source: §B.1 · pp. 15–16
  • NRG training first freezes the U-Net and VAE encoder while training ControlNet and the VFM projector, then fine-tunes the VAE decoder with ControlNet.The second phase is intended to sharpen high-frequency synthesized noise details.Source: §B.1 · p. 16

B.2.1. BYTECAMERADEPTH DATASET

ByteCameraDepth combines real depth from heterogeneous consumer sensors and scenes containing challenging materials and fine geometries. Its invalidation masks are derived from native sensor failure flags, with simulation-real pairs aligned using factory calibration.

B.2.1. BYTECAMERADEPTH DATASET

  • ByteCameraDepth includes Active Stereo, ToF, and LiDAR depth data, spanning three consumer-grade sensing technologies.The stated goal is learning physical invalidation across sensing principles rather than fitting artifacts from one hardware type.Source: §B.2.1 · p. 16
  • Dataset scenes include transparent, specular, and absorptive materials, as well as fine geometries such as wires and mesh.Examples include glass bottles, metal cans, black rubber, and fabric.Source: §B.2.1 · p. 17
  • Raw depth is log-normalized and simulation-real pairs are spatially registered using high-precision factory calibration parameters.The preprocessing also includes aligned simulation-real pairs.Source: §B.2.1 · p. 17
  • A pixel is labeled invalid when the raw sensor output is zero, NaN, or beyond the reliable range of 5.0 m.These native failure flags provide supervision for signal loss rather than manual annotation.Source: §B.2.1 · p. 17

B.2.2. NYU-DEPTH-V2 ZERO-SHOT EVALUATION

This appendix describes a zero-shot NYU-Depth-v2 evaluation, baseline implementations, depth-quality metrics, and the real-world robotic evaluation pipeline. The robotic setup spans precision manipulation and articulated-object tasks, with adapted RGB-D policies and generated demonstrations.

B.2.2. NYU-DEPTH-V2 ZERO-SHOT EVALUATION

  • NYU-Depth-v2 is used for zero-shot evaluation on its 654-image test split, representing an older Structured-Light sensor absent from training.Evaluation uses the center crop to assess transfer to unseen sensor architectures and environments.Source: §B.2.2 · p. 18, Table 14 · p. 18

B.3. Baseline Implementation

  • The baseline suite covers analytical methods and data-driven models, with five representative methods adapted and trained or tuned on ByteCameraDepth.The described analytical baselines include procedural noise and ray-tracing; data-driven baselines include translation and diffusion approaches.Source: §B.3 · p. 18, §B.3.1 · p. 18, §B.3.2 · p. 19
  • The appendix characterizes ActiveStereo as accurate for particular sensor models but computationally prohibitive and weak at generalizing to unknown sensor logic.DCL-Depth preserves global geometry but can miss sharp invalidation boundaries, while Stable-Sim2Real may hallucinate artifacts in safe regions and lacks explicit invalidation control.Source: §B.3.1 · p. 19, §B.3.2 · p. 19

B.4. Evaluation Metrics Definition

  • Evaluation metrics cover overall depth quality, sensing invalidation, and measurement inaccuracy as separate aspects.Overall metrics include MAE, RMSE, AbsRel, and threshold accuracy; invalidation metrics include IoU, recall, and F1.Source: §B.4 · p. 19, §B.4.1 · p. 19, §B.4.2 · p. 20

B.4.2. SENSING INVALIDATION (SAFETY CRITICAL)

  • Invalidation recall measures detected real invalid pixels; missing a hole can make a robot treat empty space as valid geometry, risking grasp failures or collisions.F1 is also reported to balance precision and recall under invalidation regions that typically cover less than 10% of pixels.Source: §B.4.2 · p. 20

B.4.3. MEASUREMENT INACCURACY (PRECISION)

  • Measurement-inaccuracy metrics are computed only on valid pixels, separating surface-noise fidelity from errors caused by signal loss.The text presents this separation as a way to assess the two modalities without conflating them.Source: §B.4.3 · p. 20

C.1. Simulation Environment Construction

  • The real-world evaluation uses Franka Research 3 for precision tasks and Realman RM75 for long-horizon articulation tasks, with RealSense D435 cameras for RGB-D perception.The described manipulation tasks span six tasks across the two platforms.Source: §C.1 · p. 20, Figure 9 · p. 21

C.2. Large-Scale Demonstration Collection

  • The demonstration pipeline augments ten VR demonstrations per task with MimicGen variations and whole-body-control smoothing for feasible trajectories.Three policy architectures are adapted to 16-bit depth maps using a dual-stream RGB-D encoder.Source: §C.2 · p. 21, §C.3 · p. 22

C.4. Task Descriptions

The real-world evaluation covers six manipulation tasks across two robot platforms, selected to test depth-noise challenges associated with different physical properties. The supplementary quantitative analysis examines PRISM’s design choices and hyperparameter sensitivity.

C.4. Task Descriptions

  • Six manipulation tasks span specular, transparent, deformable, and articulated challenges across two robotic platforms.The evaluation groups precision manipulation on a Franka Research 3 separately from articulated interaction on a Realman RM75.Source: §C.4 · pp. 22–23
  • On the Franka Research 3, Can-to-Plate tests holes on a reflective can, Banana-to-Plate tests measurement noise, and Stack Cubes tests cumulative geometric error.Stack Cubes requires a stable three-cube tower; the banana must be placed without slipping or falling.Source: §C.4 · p. 22
  • On the Realman RM75, Bucket-to-Plate combines translucent-body measurement inaccuracy, specular invalidation, and a swinging handle, while Toyball-to-Drawer adds a reflective handle and arm occlusion.Both tasks require interaction with articulated mechanisms.Source: §C.4 · p. 23
  • Bread-to-Microwave tests a semi-transparent grid door that creates depth artifacts, with the policy required to use the opaque handle edge and treat the window as a no-go zone.The stated success criterion is closing the door with the bread inside.Source: §C.4 · p. 23

D. Additional Quantitative Analysis

  • The additional quantitative analysis reports ablations of PRISM’s causal design, physics priors, and supervision strategy, plus sensitivity analyses of key hyperparameters.The stated aim is to assess design efficacy and robustness to hyperparameter choices.Source: D. Additional Quantitative Analysis · p. 23

D.1. Detailed Ablation Studies

The detailed ablations examine PRISM’s architecture, foundation-model physics priors, and H-PPS supervision, followed by hyperparameter sensitivity and qualitative synthesis and robot-rollout visualizations.

D.1. Detailed Ablation Studies

  • The vanilla diffusion NRG baseline has Inv-IoU 0.412 and fails to distinguish aleatoric sensor noise from deterministic physical invalidations.Table 15 defines Inv-IoU as a sensor-hole detection metric.Source: §D.1 · p. 23, Table 15 · p. 24
  • Removing SPR lowers invalidation-detection F1 from 0.975 to 0.891, indicating weaker detection without VFM-driven semantic priors.The passage relates the drop to difficulty inferring material-dependent failures on non-Lambertian surfaces.Source: §D.1 · p. 23, Table 15 · p. 24
  • Removing BND raises measurement error to MAE 0.098m from 0.076m and produces severe geometric hallucinations.The passage attributes the result to conflating distinct noise sources.Source: §D.1 · p. 23, Table 15 · p. 24
  • MoGe2-Large achieves the highest Robot SR (92.5%) among the compared VFM backbones, while larger models incur higher inference latency.The comparison covers CNNs, generic ViTs, and 3D VFMs; Robot SR averages success across six manipulation tasks.Source: §D.1 · p. 23, Table 16 · p. 24
  • With invalidation regions sparse, adding positive weight wpos = 3.0 raises recall from 48.5% with standard BCE to 71.2%.The table caption describes a progressive ablation of H-PPS components from the BCE baseline.Source: §D.1 · p. 24, Table 17 · p. 24
  • Dice Loss improves IoU by 8.3%, while PP-DHM supports fine-grained detail and dynamic decay improves Boundary-IoU by 8.8% over static mining.Boundary-IoU evaluates artifact-edge precision within 5 pixels.Source: §D.1 · p. 24, Table 17 · p. 24

D.2. Hyperparameter Sensitivity

  • The sensitivity analysis identifies wpos = 3.0 as the highest-F1 trade-off (0.975); wpos = 5.0 raises recall to 0.995 but lowers precision to 0.812.For mining, γstart = 0.25 with γend = 0.1 yields the best Boundary-IoU (0.823); the analysis reports that both smaller and larger starting ratios perform worse.Source: §D.2 · pp. 24–25, Table 18 · p. 25

E. Additional Qualitative Visualizations

  • Figures 10–14 show synthesis across varied materials, while Figures 15–20 illustrate task rollouts in which the policy treats a noisy transparent surface as an obstacle.The stated failure cases are VFM semantic ambiguity, such as mistaking a mirror for a window, and over-smoothed boundaries on structures thinner than the VFM patch resolution, such as wires under 2 mm.Source: §E.1 · p. 25, §E.2 · p. 25

F. Limitations and Future Work

The discussion identifies boundaries in PRISM’s current modeling scope and outlines extensions for more robust embodied-AI applications. It concludes that disentangling physical causality from semantic priors is key to bridging the perceptual sim-to-real gap.

F. Limitations and Future Work

  • PRISM operates as an offline data generator, and the authors propose using it to train noise-native policies that understand sensor artifacts rather than remove them.They suggest this direction could reduce reliance on auxiliary depth refinement at deployment.Source: F. Limitations and Future Work · p. 26
  • The SPR and BND modules are trained for Active Stereo, and their learned semantic-to-noise mappings do not transfer reliably to ToF or LiDAR.The authors propose lightweight per-sensor adaptation using a small number of target-sensor samples.Source: F. Limitations and Future Work · p. 26
  • PRISM models noise per frame, so independent predictions can make invalidation masks flicker across frames in dynamic scenes.The proposed extension is a flow-guided temporal state module to promote coherence.Source: F. Limitations and Future Work · p. 26
  • The current formulation focuses on intrinsic noise linked to surface appearance and scene geometry, not interaction-induced camera vibration or multi-camera crosstalk.The authors identify these extrinsic factors as a potential extension for complex multi-agent systems.Source: F. Limitations and Future Work · p. 26
  • Performance may degrade in extreme optical environments such as fog, rain, or sandstorms, where atmospheric scattering and attenuation dominate surface-reflectance-driven failures.The authors propose domain-specific adaptation of VFM priors for these environments.Source: F. Limitations and Future Work · p. 26
  • The conclusion argues that explicitly disentangling physical causality from semantic priors is key to bridging the perceptual sim-to-real gap.It calls for a move beyond monolithic noise modeling toward interpretable, physics-grounded generative simulation.Source: F. Limitations and Future Work · p. 26

Related papers

Loading 2610.29432v1…