PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Xiujian Liang, Liuyuan Na, Zehao Du, Jiacheng Xu, Weixi Song, Mingyang Sun, Xingjun Ma, Cewu Lu, Jianhua Sun

arXiv:2610.29246v1cs.ROcs.CV

TL;DR

Robot manipulation prediction must relate scene geometry to action-driven change, but prior work does not establish a common state interface between spatial understanding and action evolution. DepState-JEPA builds a depth-grounded shared RGB–depth state for spatial readout and action-conditioned prediction, and reported comparisons show gains across transfer, feature prediction, forecasting, and manipulation.

  • Problem

    Prior approaches target multimodal perception or geometric estimation but do not establish a common state interface between spatial understanding and action-driven evolution.

    Source: §1 · p. 1

  • Method

    DepState-JEPA grounds RGB and depth in metric 3D space, aggregates shared features into DepState while retaining private decoding features, and uses that state for spatial readout and action-conditioned prediction.

    Source: §1 · p. 2

  • Results

    Comparisons report stronger RGB-encoder transfer after RGB–D adaptation, feature completion, conditioned visual forecasting, and manipulation success, with ablations identifying contributions from depth, metric coordinates, and future-state matching.

    Source: §1 · p. 2, §5 · p. 9

  • Takeaways & Limitations

    Within the reported tasks, a depth-grounded shared state links spatial queries with action-driven prediction, and ablations identify contributions from geometric inputs and future-state supervision.

    Source: §1 · p. 2, §5 · p. 9

    Fixed initial private memory is a compact decoding choice, but its adequacy under large displacement and newly visible surfaces requires further study.

    Source: F LIMITATIONS · p. 45

Abstract

from arXiv · show

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 \textbf{DepState-JEPA}, a dual-stream RGB--Depth joint-embedding predictive architecture that connects spatial understanding with action-conditioned prediction through a shared latent state, \textbf{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.

1 INTRODUCTION

The paper asks how depth geometry can inform predictions of action-driven scene change, then introduces DepState-JEPA to connect spatial understanding and action-conditioned prediction through a shared RGB–depth state.

1 INTRODUCTION

  • DepState-JEPA uses a depth-grounded shared latent state to connect RGB–depth observations, spatial readout, and action-conditioned transitions.Modality-private features remain available for decoding rather than becoming separate evolving states.Source: §1 · pp. 1–2, Figure 1 · p. 2
  • Visible metric depth and camera intrinsics ground both streams in 3D, while shared–private projections and aggregation form DepState and preserve modality-specific decoding detail.Ray-conditioned readouts query local shared features without target depth; the same decoders predict future RGB and depth features using fixed initial private memory.Source: §1 · p. 2
  • Robot-state readouts provide recursive proprioception, while object-trajectory supervision and calibrated forward kinematics shape action-driven state learning.The kinematic consistency signal couples independent joint and end-effector predictions.Source: §1 · p. 2
  • Training first adapts DepState from images and video, then adds action-conditioned transitions and physical readouts while replaying RGB–D adaptation objectives.Evaluation spans spatial understanding, multi-step forecasting, and manipulation, with controlled ablations of geometry, feature sharing, and kinematic consistency.Source: §1 · p. 2
  • The reported comparisons show stronger RGB-encoder transfer after RGB–D adaptation, while within-model ablations identify contributions from depth, metric coordinates, private decoding, and future-state matching.The full pipeline supports feature prediction and manipulation.Source: §1 · p. 2

2 RELATED WORK

The related work situates DepState-JEPA among predictive representation methods, depth and 3D JEPAs, and action-conditioned robot world models. It distinguishes this work by connecting geometric representation learning to spatial completion and action-driven prediction through a shared state.

2.1 JOINT-EMBEDDING PREDICTIVE ARCHITECTURES

  • RGB-centric JEPA methods predict image or video representations, while DepState-JEPA extends this framework with an independent Depth-Video stream and measured geometry for shared-state construction.The cited approaches include I-JEPA, V-JEPA, V-JEPA 2, V-JEPA 2.1, and object-level Causal-JEPA.Source: §2.1 · p. 3

2.2 DEPTH AND 3D REPRESENTATION LEARNING FOR JEPAS

  • Depth-related JEPAs differ in whether they predict depth, use it as supervision, or encode geometric input; DepState-JEPA jointly encodes RGB and measured depth in separate streams.Its shared–private factorization and ray-conditioned readout connect geometric representation learning with spatial completion and future RGB–D feature prediction.Source: §2.2 · p. 3

2.3 ACTION-CONDITIONED WORLD MODELS FOR ROBOTICS

  • Robot world models learn action-conditioned representation changes, while DepState-JEPA adds consistency among physical predictions through calibrated forward kinematics and increment-based object-motion supervision.The kinematic constraint regularizes the state transition without guaranteeing physically correct dynamics.Source: §2.3 · p. 3, Figure 2 · p. 4
  • Concurrent work explores geometric supervision, spatial prediction, physical grounding, physics-conditioned dynamics, and joint world modeling with action generation.These directions complement DepState-JEPA’s explicit depth stream, depth-lifted positions, and selectively shared RGB–D features.Source: §2.3 · p. 3

3 METHOD

The method constructs a depth-grounded shared state from RGB–D observations, trains local feature readout without target depth, and evolves that state under robot actions to predict visual and physical outcomes.

3.1 PROBLEM SETUP AND JEPA OBJECTIVE

  • Given registered RGB, metric depth, and camera intrinsics, the method constructs DepState for local feature readout and action-conditioned evolution, using future RGB–D only as training targets.The JEPA objective predicts target representations, and the inherited dense multi-level objective supervises RGB and depth features.Source: §3.1 · p. 4

3.2 DEPTH-GROUNDED SHARED-STATE CONSTRUCTION

  • Metric depth and camera intrinsics provide each visible patch’s camera-frame position, whose learned 3D encoding is added to both RGB and depth streams.Shared and private projections separate information entering the recurrent state from modality-specific content used by visual decoders; learned queries aggregate shared features into a fixed-size state.Source: §3.2 · pp. 4–5

3.3 RAY-CONDITIONED SPATIAL PREDICTION

  • Ray-conditioned spatial readout queries local shared content using ray direction and relative time, without supplying target depth to the online query.The completion objective supervises local shared features, while dense prediction anchors modality-specific encoder features in representation space.Source: §3.3 · p. 5

3.4 ACTION-CONDITIONED STATE PREDICTION

  • An action-conditioned temporal Transformer advances DepState, and the visual readouts reuse the predicted state with fixed initial private features to predict future RGB–D features.Physical heads predict robot state and object-pose increments; direct supervision and calibrated forward kinematics couple independent joint and TCP predictions.Source: §3.4 · pp. 5–6

3.5 TRAINING AND ROLLOUT PROTOCOL

  • RGB–D adaptation learns state construction from images and video, then robot post-training adds action-conditioned transitions and physical readouts while image/video replay retains adaptation objectives.Rollouts recursively use predicted states and proprioception with fixed initial private features; Table 2 instead uses recorded proprioception.Source: §3.5 · p. 6

4 EXPERIMENTS

The experiments evaluate spatial transfer and completion, visual prediction, manipulation, and component contributions through task-specific protocols and ablations.

4.1 EXPERIMENTAL SETUP

  • The comparison includes RGB-origin and depth-related JEPAs, while within-model ablations test designated components.Complete systems retain their stated pretraining, capacity, and adaptation differences.Source: §4.1 · p. 7

4.2 SPATIAL REPRESENTATION TRANSFER AND PREDICTION

  • Frozen RGB-encoder probes assess monocular depth and segmentation transfer, while masked RGB-D completion scores features on held-out LingBot-Depth scenes.Completion uses identical masks, visible context, and calibration across methods.Source: §4.2 · p. 7, Table 1 · p. 7

4.3 TEMPORAL VISUAL PREDICTION

  • Temporal evaluation measures next-frame features and recursive forecasts at 8, 16, 24, and 32 steps using recorded actions and proprioception.Table 2 forecasts use no future images, and DepState-JEPA keeps initial private memory fixed.Source: §4.3 · p. 8, Table 2 · p. 8

4.4 EMBODIED PREDICTION AND MANIPULATION

  • Embodied evaluations cover imagined-trajectory policy learning, latent-space MPC, and frozen simulation-trained visual-goal MPC in matched simulation and real environments.The real-robot tasks are drawer and cup manipulation.Source: §4.4 · p. 8, Table 3 · p. 9

4.5 ABLATION STUDY

  • Combining depth content and metric coordinates reduces L1 from 0.408 to 0.340 and raises SR from 76.8% to 83.4%.The reported joint gains are subadditive at the means.Source: Figure 6 · p. 9, §4.5 · p. 9
  • Removing future-state matching lowers SR from 83.4% to 79.5%, while removing FK gives 82.5% with direct labels retained.Separate decoders reach L1 0.333 and SR 83.4% with added capacity.Source: §4.5 · p. 9
  • Ablations also find that removing private decoding raises L1 from 0.340 to 0.390, while mean aggregation gives L1 0.440 and SR 77.0%.Pixel and ray queries tie in feature scores; private decoding primarily benefits visual readout.Source: §4.5 · p. 9

5 CONCLUSION

DepState-JEPA connects RGB–depth geometry, shared-state evolution, and private decoding across spatial completion, conditioned forecasting, and manipulation.

5 CONCLUSION

  • RGB-D adaptation improves encoder transfer, while ablations identify contributions from geometry, local state content, and future-state supervision.The paper also reports additional control gains from FK regularization.Source: §5 · p. 9

STATEMENT ON THE USE OF AI TOOLS

The supplied material includes an AI-use statement and appendix notation references for the method's observations, spatial representation, and physical quantities.

STATEMENT ON THE USE OF AI TOOLS

  • The authors report using Codex only for manuscript drafting assistance, readability editing, and LATEX formatting and layout, under human direction and review.They state that they take responsibility for the final content.Source: STATEMENT ON THE USE OF AI TOOLS · p. 10

A NOTATION AND CONVENTIONS

A.3 ACTION CONDITIONING AND PHYSICAL QUANTITIES

  • Table 6 defines temporal, physical, and geometric notation, including joint commands, configuration, instantaneous velocity, control interval, predicted endpoint configuration and velocity, TCP pose, and object/part body increments.It also specifies independent joint-configuration and TCP-pose heads, an object/part integration origin, and an integrated next object/part pose.Source: Table 6 · p. 14, §A.3 · p. 15

A.4 OBJECTIVES AND TRAINING PARAMETERS

The supplied passages identify Table 7's notation scope and describe the focus of the representation-architecture section.

A.4 OBJECTIVES AND TRAINING PARAMETERS

  • Table 7 lists objectives and schedule symbols used in the main method.Source: Table 7 · p. 15

B REPRESENTATION ARCHITECTURE AND SPATIAL LEARNING

B.1 DATA INTERFACES, GEOMETRY, AND MASKING

The data protocol pairs registered RGB-D observations with metric geometry and synchronized masking, then builds and predicts shared states while retaining private features for decoding. Momentum targets supervise completion and prediction, with target features stopped from receiving gradients.

B.1 DATA INTERFACES, GEOMETRY, AND MASKING

  • A fixed calibrated external RGB-D camera and embodiment-specific post-training define the primary robot protocol; wrist-camera-only trajectories are outside its scope.Clips do not cross trajectory boundaries, and robot states are encoded from the single RGB-D image at each control timestamp.Source: §B.1 · p. 15
  • Metric depth and camera intrinsics place valid patches in camera-coordinate 3D, while invalid geometry receives an unknown embedding and cannot supply geometric targets.Depth patches with fewer than 50% valid pixels are excluded from aggregation; synchronized masks remove masked embeddings before encoder attention.Source: §B.1 · p. 16

B.2 BACKBONES, STATE CONSTRUCTION, AND PREDICTION HEADS

  • The state architecture uses metric position encoding and ray/time-conditioned shared-feature queries, then decodes modality details through private features.The same spatial and modality decoders serve spatial denoising and future prediction; future prediction does not read future private features.Source: §B.2 · pp. 17–18, Table 8 · p. 17

B.3 MOMENTUM TARGETS AND GRADIENT BOUNDARIES

  • A momentum teacher supplies stopped completion, dense-feature, and future-state targets, while online alignment and physical-consistency predictions retain gradients.Teacher and online slots keep corresponding query indices rather than being rematched at each timestep.Source: §B.3 · p. 18

C TRAINING LOSSES AND PHYSICAL CONSISTENCY

The objective combines representation learning with direct physical supervision and kinematic consistency. Training adapts RGB-D representations first, then post-trains action-conditioned prediction while replaying adaptation data.

C. TRAINING LOSSES AND PHYSICAL CONSISTENCY

  • Direct physical supervision covers endpoint joint configurations, TCP poses, object poses, and instantaneous endpoint joint velocities when valid annotations are available.Object trajectories are supervised only where a valid initial pose and corresponding endpoint labels are available.Source: §C.1 · p. 19, §C.2 · p. 20

C.1 COMPLETE LOSS DEFINITIONS

  • The robot objective combines temporal matching, feature-factorization regularization, direct physical supervision, and a ramped forward-kinematics consistency term.FK consistency is zero-weighted for the first 2,000 robot updates, reaches full weight at 5,000, and remains fixed thereafter.Source: §C.1 · pp. 18–19

C.2 PHYSICAL READOUTS, UNITS, AND KINEMATIC CONSISTENCY

  • Calibrated differentiable forward kinematics couples independently supervised joint and TCP predictions through the same predicted state and known robot geometry.Consistency requires compatible base and tool frames, joint order, units, zero offsets, and timestamps; it couples joint and TCP predictions, not TCP contact with the object.Source: §C.2 · p. 20

D.1 OPTIMIZATION PROTOCOL

  • RGB-D adaptation precedes robot post-training, which retains adaptation objectives through replay while training action-conditioned transitions and physical readouts.The rollout horizon is one for the first 2,000 robot updates and three thereafter; recorded future commands condition rollouts, but future measurements do not replace recursive predictions.Source: §D.1 · pp. 20–21, Table 9 · p. 21

D.2 TWO-STAGE TRAINING ALGORITHM

The training algorithm separates RGB-D representation adaptation from robot post-training, where action-conditioned states are recursively predicted and supervised. Inference uses the online stack, and the evaluation passages specify frozen-transfer datasets and protocols.

D.2 TWO-STAGE TRAINING ALGORITHM

  • Two-stage training first learns RGB-D representations, then adds a temporal predictor and physical heads for robot post-training with periodic representation replay.Robot post-training uses a one-step horizon for the first 2,000 robot updates and a three-step horizon thereafter.Source: §D.2 · p. 22

D.3 OPEN-LOOP PREDICTION AND TRAINING ALGORITHM

  • Each open-loop step predicts a next state, endpoint proprioception, and object increment, then queries future RGB-D features using fixed initial private memory.Recursive predictions feed the next step; future observations and labels are used only during training for targets and losses.Source: §D.3 · pp. 22–23

D.4 INFERENCE INTERFACES

  • Inference uses the online encoder, state constructor, predictors, and heads; RGB-only probes use the adapted RGB encoder without depth, metric position encoding, or DepState.Absolute object trajectories require an initial object pose, although relative increments do not.Source: §D.4 · p. 23

E.1 DATASETS FOR FROZEN SPATIAL TRANSFER

  • NYUv2 depth probes use the same 654 labeled test images across methods, while VOC12 segmentation probes use the official training and validation splits.NYUv2 test depth is used only for scoring, not scale alignment or model selection; VOC12 results are reported on the validation split, not the hidden test set.Source: §E.1 · p. 24

E.2 METRICS FOR FROZEN SPATIAL TRANSFER

The frozen-transfer metrics define depth and segmentation endpoints with distinct aggregation procedures. Lower depth errors and higher threshold accuracy or segmentation scores indicate better performance.

E.2 METRICS FOR FROZEN SPATIAL TRANSFER

  • Depth scores compare predicted and ground-truth depth over valid pixels, with lower errors and higher δ1 indicating better performance.Predictions are restored to the original label grid and scored using the shared mask and declared depth range.Source: §E.2 · p. 24
  • Depth metrics are computed per image and averaged equally across the 654 NYUv2 test images; RMSE is the primary depth endpoint.AbsRel and log10 are dimensionless, RMSE is reported in meters, and δ1 is a percentage.Source: §E.2 · p. 24
  • VOC12 mIoU and mean class accuracy are percentages computed from the full validation-set confusion matrix, excluding void pixels and including background.mIoU is the primary segmentation endpoint; mAcc gives equal weight to per-class recall.Source: §E.2 · pp. 24–25

E.3 FROZEN SPATIAL TRANSFER: READOUTS AND TRAINING

This section specifies protocols spanning frozen spatial-transfer probes, RGB-D feature completion, action-conditioned visual prediction, and embodied control, including control learning and planning, sim-to-real training and deployment, metrics, and task success criteria. It describes shared scoring protocols alongside method-specific representations and adaptations, but provides no comparative result values.

E.3 FROZEN SPATIAL TRANSFER: READOUTS AND TRAINING

  • Frozen spatial transfer evaluates encoder features with common supervised probes, while representation training and pretraining remain method-specific.Encoders stay frozen during depth and segmentation probe fitting; NYUv2 depth and VOC labels train only their respective heads. JEPADepth’s NYUv2-self-supervised context encoder is evaluated after removing its native depth-estimation modules, and its VOC12 result tests cross-dataset feature transfer rather than a native segmentation model.Source: §E.3 · pp. 25–26

E.4 RGB-D SPATIAL COMPLETION

  • RGB-D completion predicts shared frozen RGB and depth reference features from visible, calibrated context under synchronized masks.Approximately 50% of patch-aligned image blocks are withheld from both modalities before online feature extraction and state aggregation; queries use target pixels or camera rays, never hidden content. Completion uses equal-weight masked-region L1 losses, and the benchmark measures missing-content inference rather than sensor-noise removal.Source: §E.4 · p. 28

E.5 TEMPORAL VISUAL PREDICTION

  • Temporal prediction compares action-conditioned adaptations on common robot episodes and horizons, scoring predicted RGB features after mapping them to a shared frozen V-JEPA 2.1 reference space.The evaluation includes V-JEPA 2-AC, V-JEPA 2.1-AC, 3D-JEPA-AC, Le MuMo JEPA-AC, and DepState-JEPA; action-free encoders alone are not temporal predictors. Each method’s frozen linear readout is fit on paired observed training-frame features and then reused for predicted features at every horizon.Source: §E.5 · pp. 29, 31

E.6 EMBODIED PREDICTION AND MANIPULATION

  • Embodied evaluation uses separate agents on 20 DMControl and 50 Meta-World tasks, with shared task sets, camera settings, and reset schedules within each control track.Methods receive common camera views, goals when required, and current proprioception; multimodal methods also receive aligned metric depth and calibration. Controllers do not receive privileged simulator state, and proprioception is predicted rather than read from future environment states during imagined rollouts.Source: §E.6 · p. 32
  • Control performance is evaluated over three seeds on ten episodes per task, using final-budget checkpoints and separate suite metrics: DMControl episode return and Meta-World final-step success.Suite means are averaged equally across tasks; evaluation transitions are excluded from training and checkpoint selection.Source: §E.6 · p. 32

E.6.1 POLICY LEARNING THROUGH IMAGINED TRAJECTORIES

  • DreamerV3-style actor–critic learning uses 15-step imagined trajectories initialized from replay states, with observed next-frame encodings supervising latent prediction.The actor’s gradients do not update the world model, and evaluation uses its mean action without exploration noise or candidate search.Source: §E.6.1 · pp. 32–33

E.6.2 RECEDING-HORIZON PLANNING IN LATENT SPACE

  • The planning track follows TD-MPC2 (Hansen et al., 2024), adapting a transition, reward and value heads, and a stochastic policy prior from replay before receding-horizon MPPI search.All representations share the control recipe and search budget; candidates use predicted proprioception, not future images, recorded future states, or simulator signals.Source: §E.6.2 · pp. 33–34

E.6.3 SIM-TO-REAL MANIPULATION TRANSFER

  • For sim-to-real training, drawer and cup environments match robot geometry, workspace, action conventions, and camera calibration; a fixed pretrained PRISM model augments simulated RGB-depth observations.PRISM synthesizes missing returns and perturbs valid depth; it is not applied to measured RGB-D at deployment.Source: §E.6.3 · p. 34
  • At deployment, frozen models use shared ordered visual subgoals and observation-based stage criteria, then replan after each action using an eight-step CEM configuration.Goal observations specify tasks but provide no action supervision; no task-specific real-world fine-tuning is used. The two full tasks require an ordered drawer open-close cycle and a complete cup grasp-lift-place sequence.Source: §E.6.3 · pp. 34–35
  • Real-robot evaluation reports full-task success separately for drawer and cup tasks, with 30 trials per trained seed and paired initial configurations across methods.Timeouts, drops, and execution failures count as failures; Table 11 reports mean ± sample standard deviation across seeds for six atomic and full tasks, while Table 12 gives success counts and denominators.Source: §E.6.3 · p. 35, Table 11 · p. 36, Table 12 · p. 36

E.7 ABLATION RESULTS, PROTOCOLS, AND STATISTICAL ANALYSIS

The ablations compare design choices against the full Ex.16 configuration, with controlled training and evaluation protocols. They find contributions from depth content and metric coordinates, distinguish visual-decoding effects from control, and report temporal and physical-supervision trade-offs.

E.7 ABLATION RESULTS, PROTOCOLS, AND STATISTICAL ANALYSIS

  • Ablations use Ex.16 as the common reference, with three-seed means and fixed training and control budgets; Figure 6 connects independent interventions, not cumulative steps.Reported standard deviations are across training seeds, not confidence intervals over scenes or trials.Source: §E.7 · p. 39

E.7.4 SEED-LEVEL AND TASK-LEVEL RESULTS

  • The SR aggregate weights 50 Meta-World tasks and two real tasks, so simulated control accounts for 50/52 of its task weight.L1/Cos task means instead cover RGB-D completion and two held-out robot forecasting tasks; their forecasting conditioning differs from Table 2.Source: §E.7 · p. 39, §E.7.4 · p. 42

E.7.1 GEOMETRY-GROUNDED STATE CONSTRUCTION

  • Ex.16 reaches L1 0.340 and SR 83.4%; depth content and metric coordinates each improve both measures, with combined gains below the sum of individual gains.The comparisons use Ex.1 as the baseline for individual factors and report complementary contributions without establishing super-additivity.Source: Table 13 · p. 40, §E.7.1 · p. 40

E.7.2 SELECTIVE SHARING AND SPATIAL READOUT

  • Mean aggregation (Ex.6) yields L1 0.440, Cos 0.799, and SR 77.0%; Ex.7 and Ex.16 tie in L1/Cos at reported precision.Removing private decoding primarily affects visual prediction (L1 0.390), while SR remains 83.2%; ray and pixel queries have comparable feature accuracy under this camera protocol.Source: Table 14 · p. 41, §E.7.2 · p. 41

E.7.3 TEMPORAL PREDICTION AND PHYSICAL GROUNDING

  • Removing future-state matching lowers SR to 79.5%, while removing future-feature matching worsens L1 to 0.392; FK raises aggregate SR from 82.5% to 83.4%.Removing FK reduces Drawer/Cup full-task success from 81.1/78.9% to 77.8/74.4%; joint and velocity supervision remain active.Source: Table 15 · p. 42, §E.7.3 · p. 42, Table 18 · p. 44

F LIMITATIONS

The demonstrated setting is bounded to fixed-camera, embodiment-specific drawer and cup tasks. The paper also cautions that feature accuracy does not establish physical trajectory accuracy, and that memory and kinematic design choices leave open important cases.

F LIMITATIONS

  • Real-world evaluation uses a calibrated fixed external camera and embodiment-specific training for drawer and cup tasks; moving-camera prediction and zero-shot embodiment transfer remain outside scope.Broader object and workspace changes are also outside the demonstrated scope.Source: F LIMITATIONS · p. 45
  • Whole-image feature accuracy alone does not establish action sensitivity in moving regions or accurate physical trajectories.The shared/private projections encourage predictive-role division but do not provide an identifiable semantic decomposition.Source: F LIMITATIONS · p. 45
  • Fixed initial private memory may be inadequate under large displacement or newly visible surfaces; FK and valid SE(3) parameterization do not guarantee contact dynamics or collision-free control.The paper identifies further study of private-memory adequacy under these conditions.Source: F LIMITATIONS · p. 45

Related papers

Loading 2610.29246v1…