acceptodds
Under review as a conference paper at ICLR 2027

GeoWRA: Geometric Latent World Residual Action Models

Abstract

World models and vision-language-action (VLA) models provide complementary capabilities for robotic manipulation through predictive modeling and multimodal action generation. However, effectively integrating world models and VLAs into a unified architecture remains a key challenge, as existing approaches are structurally decoupled, lacking intrinsic interaction between world prediction and action generation. To address this challenge, we propose Geometric Latent World Residual Action Models (GeoWRA), a unified framework that couples latent world prediction with action generation. First, we establish a bidirectionally coupled architecture in which the VLA's latent actions condition future prediction, while predicted futures in turn guide action generation through bounded world residual feedback. Second, to capture the relational structure of embodied scenarios, we introduce a Euclidean-hyperbolic product representation that jointly models future features and their relationships by embedding predicted future states in hyperbolic space, providing a geometric inductive bias for learning task-relevant predictive representations. Third, we construct a joint action and geometric world objective that integrates action supervision with future prediction under hyperbolic geometric regularization, enabling geometrically consistent future representations to guide action generation. Extensive evaluations in both simulation and real-world settings show that GeoWRA delivers competitive performance across a diverse range of robotic manipulation tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Model, Robotic, Geometric

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.67 over 3 reviews, above 99% of the others.
Reviewer_D9xVOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GeoWRA, which aims to let the predicted future in a VLA directly guide action generation, rather than serving only as an auxiliary training signal. A lightweight predictor takes features of the current observation from a frozen V-JEPA encoder and, conditioned on the VLA's action representations, predicts future V-JEPA features at several horizons. These predictions are fed back into the action tokens through a bounded residual, so the action expert is conditioned on the predicted future at every denoising step. A distinctive design choice is to embed part of the prediction in hyperbolic space, which the authors use to model relations among nested image regions and across time. The method is built on π0.5 and a Qwen-based VLA, and is evaluated on LIBERO, LIBERO-Plus, SimplerEnv and six real-world tasks.

Critical strengths
  1. Predicted futures directly guide action generation. Prior latent world action models use future prediction only as an auxiliary loss. Here, the predictor is conditioned on the current noisy action, and its prediction is fed back into the action tokens at every denoising step. The feedback is bounded and zero-initialized, so the module can be added to an existing VLA without disturbing its original behavior. It is also lightweight and has been applied to several different VLA backbones.
  2. Structural supervision on the predicted future. Beyond aligning individual future features, the paper also supervises the relations among predicted regions and across time, using a hyperbolic embedding. This brings hyperbolic representation learning into world action models, and gives the predictor a structural constraint that plain feature alignment does not provide.
  3. Evidence that the policy relies on the predicted content. The real-world tasks are chosen to require anticipating the consequences of actions, such as intercepting a moving object or keeping an unstable stack intact, and the method improves on all of them. The intervention experiments go a step further: replacing the predicted future with current features, or with futures from other trajectories, clearly lowers success. This suggests that the policy actually uses the predicted content, rather than only benefiting from extra parameters.
Critical weaknesses
  1. The motivation for hyperbolic geometry remains unclear. The spatial hierarchy consists of shallow, fixed grids of image regions. Although the Euclidean–Euclidean comparison supports a benefit from hyperbolic supervision, it is reported only on SimplerEnv. The paper also does not examine whether the embeddings capture the intended hierarchy, leaving the proposed geometric explanation insufficiently validated
  2. The benefit of future prediction is not fully separated from that of additional V-JEPA features. The StarVLA baseline lacks this feature pathway. Although inference-time substitutions show that the trained policy uses its predicted inputs, they may introduce a distribution mismatch. A baseline trained to use current V-JEPA features through the same adapter, without future prediction, would better isolate the contribution of predicting the future.
Questions for authors
  1. Could you elaborate on why hyperbolic geometry is particularly suitable for the shallow, fixed hierarchy of image regions used in this method? The Euclidean–Euclidean comparison provides evidence of an empirical benefit, but the connection between this benefit and the proposed hierarchical inductive bias remains unclear. Is there evidence that the hyperbolic representation better captures the intended nesting relationships, rather than primarily changing how feature prediction errors are weighted during training?
  2. Could you clarify how the contribution of future prediction can be separated from the benefit of introducing additional V-JEPA features? The feedback-content interventions show that the trained policy is sensitive to its predicted inputs, but inference-time substitutions may introduce a feature-distribution mismatch. Have you considered training a baseline that uses current V-JEPA features through the same residual adapter, with comparable capacity and training budget? Such a comparison would help establish what future prediction adds beyond the additional visual representation.
  3. Could you test whether the model predicts different futures for different actions from the same starting state? For example, in simulation, you could execute a few different action sequences from the same scene and compare the predicted features with the features of the actual outcomes. It would be useful to check whether providing the correct action gives a more accurate prediction than providing a different action. This would build on the existing ablations and help show whether the predictor has learned how actions change the scene, beyond what it can infer from the observation and task alone.
3 comments
AuthorsOpenReview ↗

Motivation for hyperbolic geometry, requested current-feature control, and same-state action test

Thank you for the precise questions and for recognizing that "predicted futures directly guide action generation" and that "the policy actually uses the predicted content." We have revised the paper accordingly, with all changes highlighted in blue.

W1 & Q1: Motivation and Evidence for Hyperbolic Geometry.

For the motivation, we do not rely on hyperbolic capacity, which three nested grids (21 regions) do not require. The fixed nesting defines parent–child and cross-horizon pairs that must be matched jointly, and a tree-like distance suits the nested grids while the same metric also measures the cross-horizon edges (Sec. 3.3 of our revised paper). The learned representations capture this structure. On SimplerEnv, the Spearman correlation between predicted and teacher relation distances (rank-based, as training metrics differ) reaches 0.82/0.79/0.72 for GeoWRA on spatial/temporal/nonadjacent pairs versus 0.73/0.68/0.60 for E–E, trained with relation losses on the same pairs (Table 5 in our revised main paper). The gain also holds on nonadjacent pairs, which the relation loss does not directly supervise, suggesting that it extends beyond these edges. GeoWRA without relation matching (0.76) still exceeds E–E with it (0.73), so the gain is not due to relation matching alone. Figure 4 in our revised main paper visualizes the learned nesting.

For the reweighting question, we add a conformal-weighted E–E that scales its flat losses by the conformal factor λc\lambda_c, a position-dependent reweighting without curvature. We also extend E–E (the zero-curvature limit with identical banks, pairs, and budget) beyond SimplerEnv to all four settings. As shown in Table R4, both Euclidean variants remain below GeoWRA in all four settings, so this reweighting recovers part, but not all, of the gain and does not fully explain it.

Table R4 (Table 14 in our revised appendix). Geometry controls: success rates (%).

MethodLIBEROLIBERO-PlusGoogleWidowX
E–E97.685.366.868.8
Conformal-weighted E–E97.785.767.970.1
GeoWRA98.086.269.171.9

W2 & Q2: Separating Future Prediction from Additional V-JEPA Features.

We train the suggested current-feature control, which uses current V-JEPA features through the same residual adapter without future prediction, with comparable capacity, data, and training budget. As shown in Table R5, this additional visual representation improves StarVLA by 7.2 points. With the same adapter, future prediction adds a further 6.9 points. Because the control is trained and evaluated with current features, it also avoids the distribution mismatch you raise. The 6.9-point gap is thus free of it, and the control reaches 63.6% versus 61.8% for the inference-time substitution of Table 11 in our revised appendix (Table 9 in the original submission).

Table R5 (from Table 4 in our revised main paper). Contribution of each component on SimplerEnv.

MethodV-JEPA featuresFuture predictionOverall (%)Gain
StarVLA✗✗56.4–
Current adapter (trained)✓✗63.6+7.2
GeoWRA✓✓70.5+6.9

Q3: Different Actions from the Same Starting State.

Thank you for this constructive suggestion. From each of 120 initial scenes (30 per task) with a fixed instruction, we execute four distinct seven-step action sequences, so only the action differs. We compute the NRMSE between predicted and actual-outcome features at τ=1\tau=1, conditioning on the generating action (correct) or, averaged, on the other three (incorrect). Identification accuracy is the fraction of the 480 outcomes whose generating action gives the lowest NRMSE, and a predictor that ignores the action stays at chance (25%). As shown in Table R6, on all four tasks, the correct action gives a more accurate prediction than a different action, with 22.3% lower NRMSE and 89.0% accuracy. In this setting, the predictor has thus learned how actions change the scene, beyond what it can infer from the observation and task alone.

Table R6 (Table 16 in our revised appendix). Same-state action-conditioned prediction on SimplerEnv–Google (last-row reduction from the mean NRMSE).

TaskCorrect ↓Incorrect ↓Reduction ↑Accuracy ↑
Pick up a Coke can0.39890.535425.5%97.5%
Move an object near another0.36150.447419.2%80.8%
Open/close a drawer0.38150.489422.0%85.8%
Place an apple in the top drawer0.40160.513521.8%91.7%
Macro-average0.38590.496422.3%89.0%

We hope these responses address your concerns and are glad to answer further questions.

Thank you for the detailed response and the additional experiments. The trained current-feature control directly addresses my concern about separating the contribution of future prediction from that of additional V-JEPA features. The same-state, different-action experiment also provides convincing evidence that the predictor uses action information to distinguish future outcomes in the evaluated setting.

The expanded Euclidean controls and relation-distance analysis strengthen the empirical justification for the hyperbolic formulation. While the specific explanation in terms of a hierarchical inductive bias remains suggestive rather than conclusive, this is now a limited concern rather than a major weakness.

Overall, the additional experiments substantially address my original concerns.

AuthorsOpenReview ↗

Thank you for recognizing our work and raising the score to 4

Thank you for raising your score to 4 and for the careful follow-up. Your suggestions on the trained current-feature control, the same-state action test, and the expanded geometry controls directly shaped the new experiments and substantially improved the paper. We sincerely appreciate your time and constructive feedback throughout the review process.

Reviewer_hKxeOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoWRA, a framework that integrates action-conditioned latent world prediction with VLA action generation through bounded residual feedback. It further introduces Euclidean-hyperbolic representations and geometric supervision to capture spatial-temporal relationships in predicted future states.

Critical strengths

The paper demonstrates consistent performance improvements across multiple simulation benchmarks and real-world robotic manipulation tasks, with extensive ablation studies supporting the effectiveness of bidirectional coupling, geometric supervision and bounded residual feedback.

Critical weaknesses
  • Limited contribution. The proposed framework primarily combines existing methods including latent world prediction, attention-based feature fusion, residual feedback, and hyperbolic representation learning. More insights and fundamental contributions beyond combining these established components is not sufficiently clear. The authors should clarify the fundamental contributions compared with prior methods.

  • Insufficient analysis of the learned hyperbolic representations. The paper motivated by by the property of hyperbolic geometry that captures hierarchical and spatio-temporal relationships. However, it remains unclear whether the learned representations actually show these properties. Even if the authors provide Euclidean baselines and ablations, improvements in downstream success rates cannot fully justify the representational benefits of hyperbolic geometry.

Questions for authors
  • What is the key contribution or distinctions of GeoWRA beyond integrating existing world modeling, residual feedback and geometric representation techniques?

  • Can the authors provide visualizations or quantitative analysis demonstrating the learned hyperbolic representations and their meaningful hierarchical and spatial-temporal structures?

3 comments
AuthorsOpenReview ↗

Fundamental contribution and requested analysis of the learned hyperbolic representations

Thank you for the constructive comments and for recognizing the "consistent performance improvements" across simulation and real-world tasks and the "extensive ablation studies" supporting our design. We have revised the paper accordingly, with all changes highlighted in blue.

W1 & Q1: Key Contribution beyond Integrating Existing Techniques.

For the core contribution, the key idea of GeoWRA is bidirectional predictive coupling within action generation (Sec. 3.2). From action to world, the world predictor conditions latent future prediction on the latent-action tokens qtq_t and the evolving action aτa_\tau at flow time τ\tau. From world to action, a residual adapter reads the predicted futures using action tokens as queries and returns a bounded world residual at the same flow step, so the updated action conditions the next prediction. The fundamental difference from prior paradigms (Sec. 2.1) lies in where prediction acts. Planning-based models (V-JEPA 2-AC, LeWorldModel) use it before generation, selecting actions by searching over predicted futures. Joint modeling (WorldVLA) and predictive supervision (VLA-JEPA) use it mainly in training to shape shared parameters or representations, rather than condition the action at inference. Representation transfer (JEPA-VLA) injects current V-JEPA features without predicting the future. GeoWRA instead makes action-conditioned future prediction part of action generation, while retaining the underlying action expert and its original generation schedule. The hyperbolic factor structures the futures fed back to the action, supervising their nested spatial and cross-horizon relations.

For the empirical validation, our experiments support each part of this design. Both directions contribute: removing action conditioning or predictive feedback costs 5.70/8.55 points on SimplerEnv, and the Euclidean-only variant costs 4.05 points (Figure 3 in our revised main paper). The predictor depends on the action: with the observation fixed, it identifies the generating action for 89.0% of outcomes (chance 25%, Table 16 in our revised appendix). The policy uses the predicted content: replacing it with current features or mismatched futures lowers success by 8.70/11.75 points (Table 11 in our revised appendix). Notably, a current-feature control akin to representation transfer, with the same interface, trails GeoWRA by 6.9 points (in our revised Table 4).

W2 & Q2: Analysis of the Learned Hyperbolic Representations.

Thank you for this suggestion. Beyond success rates, we further analyze whether the learned hyperbolic representations capture the hierarchical spatial and temporal relations they are supervised for (Secs. 3.3–3.4), through a quantitative analysis and a visualization, both added to our revised paper (Table 5 and Figure 4 in Sec. 4.2, App. B.8).

For the quantitative analysis, teacher relation distances are distances between the frozen V-JEPA targets of paired regions in the recorded future (App. C.2), reflecting how close each child stays to its parent and how much each region changes over time. A high Spearman correlation with them means the predictions preserve this relational ordering, not just individual points. Spatial pairs link parent and child regions, temporal pairs the same region at adjacent horizons, and nonadjacent pairs it at non-adjacent horizons, not directly supervised by the relation loss. All variants use the tangent-readout Euclidean distance that E–E is trained on. As GeoWRA is trained on hyperbolic distances, we compare rank order rather than absolute errors. As shown in Table R3, GeoWRA reaches 0.82/0.79/0.72. Its gain over E–E (0.09/0.11/0.12) also holds on nonadjacent pairs, suggesting structure beyond the supervised edges. Even without relation matching, GeoWRA (0.76) exceeds E–E with it (0.73), so the gain is not due to relation matching alone.

Table R3 (Table 5 in our revised main paper). Spearman correlation of relation distances on SimplerEnv (↑).

RepresentationSpatialTemporalNonadjacent
E–E0.730.680.60
GeoWRA w/o relation matching0.760.710.64
GeoWRA0.820.790.72

For the visualization, Figure 4(a) projects GeoWRA's hyperbolic predictions into a 2-D Poincaré disk, with global, coarse, and fine regions (L0/L1/L2) linked by parent–child geodesics. Only the pairing is given by the grid, while positions are learned and embedded without hierarchy labels. Children of the same parent form coherent bundles, so the learned geometry reflects the containment hierarchy. Both analyses indicate that the hyperbolic representations learn meaningful hierarchical and spatial-temporal structure.

We hope these responses address your concerns and are glad to answer further questions.

Thank you for the response and clarification.

For W2/Q2, the new quantitative analysis (Table 5) and the Poincaré-disk visualization (Figure 4) address my concern well, especially Figure 4 was what I wanted to see.

For W1/Q1, joint action-to-world and world-to-action coupling unlike recent works (e.g. LingBot-VA [1]) being a supportive design for geometric representation rather than long-horizon rollout, and their application and analysis made the contribution clearer and sufficiently distinct.

Together, I am convinced to raise my scores to clear accept.

[1] Native Video-Action Pretraining for Generalizable Robot Control, Arxiv 2607.08639

AuthorsOpenReview ↗

Thank you for acknowledging our contributions and raising the score

Thank you for the follow-up and for raising your score. We are glad that the quantitative analysis in Table 5 and the Poincaré-disk visualization in Figure 4 address your concern, and that the clarified contribution is now clear and sufficiently distinct. We sincerely appreciate your time and constructive feedback, which substantially improved the paper.

Reviewer_6MzvOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GeoWRA adds an action-conditioned future-feature predictor to a VLA. At each flow step, the predictor takes the current visual features, latent-action tokens, and noisy action sample. Its predictions are fed back to the action expert through a bounded residual on the conditioning tokens. The model learns Euclidean and hyperbolic representations using frozen V-JEPA features as future targets, together with action supervision. The paper evaluates the method on three simulation benchmarks and six real-robot tasks, with ablations on action conditioning, predictive feedback, and the geometric losses.

Critical strengths
  1. The predicted features are used during action generation. The part I find most useful is how prediction is connected to the action expert. The predictor reads the action sample being generated and uses the predicted future features to condition the action expert at the same flow step. Zero initialization and RMS-based bounds control the change to the conditioning tokens. The model also retains the original number of generation steps. Removing this feedback reduces the SimplerEnv average by 8.55 percentage points in Figure 3, which supports the value of the feedback path.

  2. The experiments support the main design choices. The improvement over the matched StarVLA baseline, from 56.4% to 70.5% on SimplerEnv, is convincing. Table 5 is an important control: it retains both prediction banks and their sizes, and the Euclidean–Euclidean version reaches 67.8% versus 70.5% for the full model. This gives evidence for an additional benefit from the hyperbolic metric. Tables 9 and 11 also examine whether the policy uses the feedback content and whether action conditioning improves prediction. Together, these comparisons provide reasonable support for the proposed design.

Critical weaknesses
  1. The contribution of future prediction is not fully isolated.

GeoWRA adds both V-JEPA features and a trainable path from those features to the action expert. I am not yet clear how much of the gain comes from predicting the future rather than from this additional visual input. Table 9 replaces the feedback only at inference, after the policy has been trained to use predicted futures. The drop shows that the policy depends on its feedback, but the changed input distribution can also contribute.

A more direct control would train a current-feature adapter with the same encoder and residual interface, using comparable capacity, data, and training updates. The existing results support the full method, but do not fully separate these sources of improvement.

  1. The π0.5 integration and temporal setup need more detail.

The Qwen adapter dimensions are specified, but the implementation in π0.5 is still unclear to me. Which representations provide q_t and C_t, and which tensor receives the residual?

The temporal setup also matters: how are the eight target frames sampled, what physical times do the three prediction horizons represent, and how do they align with the action chunk? These choices determine what the predictor is learning to forecast. The training-data counts, batch sizes, and update counts should also be reported explicitly. Stating that settings are matched does not provide enough information to reproduce the results.

Questions for authors

How does a separately trained current-feature adapter perform with the same V-JEPA encoder and residual interface? Please keep the capacity, data, and update budget comparable. This is the comparison I would find most useful for assessing the added value of future prediction. Please clarify the π0.5 implementation and temporal alignment. What are the sources of q_t and C_t, and where is the residual injected? Which backbone is used in the real-robot experiments? How do the eight target frames, three prediction horizons, and action chunk align in physical time? Please also provide the actual training-data counts and training budgets. What are the end-to-end latencies per action chunk for the base policy and GeoWRA, on the same hardware and with the same number of generation steps? Please include V-JEPA encoding. The reported 1.28 GFLOPs excludes it, so the current estimate is not enough to assess the overhead in closed-loop control.

2 comments
AuthorsOpenReview ↗

Requested current-feature control and π0.5 integration (Part 1/2)

Thank you for the constructive comments and for recognizing that "the predicted features are used during action generation" and that "the experiments support the main design choices." We have revised the paper accordingly, with all changes highlighted in blue.

W1 & Q1: Isolating the Contribution of Future Prediction.

Following your suggestion, we train a current-feature control with the same frozen V-JEPA encoder and residual interface, keeping capacity, data, and training steps comparable to GeoWRA. The adapter keeps CtC_t as queries but reads current V-JEPA tokens instead of predicted futures, while aτa_\tau and τ\tau enter the action expert as in GeoWRA. It is trained and evaluated with current features, without future targets.

Table R1 (Table 4 in our revised main paper). Current-feature control on SimplerEnv: success rates (%).

MethodGoogleWidowXOverall
StarVLA57.655.256.4
Current features (inference only)62.161.561.8
Current adapter (trained)63.963.263.6
GeoWRA69.171.970.5

First, the trained control improves StarVLA by 7.2 points, showing that the V-JEPA features help as feature enhancement. Second, under the same encoder and interface, GeoWRA further exceeds this control by 6.9 points, indicating that action-conditioned future prediction improves action generation beyond the additional V-JEPA features. The control avoids the distribution mismatch you point out, so the 6.9-point gap is free of it. It reaches 63.6% versus 61.8% for the inference-time substitution of Table 11 in our revised appendix (Table 9 in the original submission).

W2 & Q2: π0.5 Integration, Temporal Setup, and Training Details.

Thank you for these questions. We have added the details to our revised paper (Sec. 3.2, App. C).

For π0.5 integration, the observation prefix is encoded once per action-generation call, with its keys and values cached across flow steps, and serves as CtC_t. qtq_t is a 64-query, width-512 readout of the final prefix states with projected proprioception, fixed within a call (Eq. 2). To preserve prefix-cache reuse, we add the residual to the expert's output velocity, with queries from its final hidden states Xt,τX_{t,\tau} and residual bounds based on the uncorrected velocity:

vΘ=vθ(aτ,τ;Ct,st)+Δvt,τ,Δvt,τ=Bτ(Rt,τ;vθ).v_\Theta=v_\theta(a_\tau,\tau;C_t,s_t)+\Delta v_{t,\tau},\quad \Delta v_{t,\tau}=\mathcal B_\tau(R_{t,\tau};v_\theta).

The adapter reads the short- and long-horizon predictions with Xt,τX_{t,\tau} as queries, and its output projection gives six motion-coordinate corrections. Bτ\mathcal B_\tau applies separate RMS caps to translational and rotational corrections, relative to the corresponding uncorrected velocity scales, followed by the flow-time gain, and gripper and padded dimensions receive zero correction. All π0.5 results in the original submission use this implementation, the velocity-level counterpart of the default Qwen form in Eqs. 3–5 (App. A.1, C.1). The predicted future corrects the velocity and reaches the expert via the next step's aτa_\tau. Removing it costs 1.2/2.7 points on LIBERO/LIBERO-Plus (Figure 3). The objectives, prefix cache, and original flow-step schedule are preserved. With all other parameters fixed, zero-initializing the residual output projection recovers the uncorrected velocity of the same backbone, and an analogous velocity-level perturbation bound follows under the assumptions of Cor. A.2′.

AuthorsOpenReview ↗

Temporal setup, training details, and end-to-end latency (Part 2/2)

Thank you again for the detailed questions on implementation and reproducibility. We continue our response below.

W2 & Q2 (continued): Temporal Setup and Training Details.

For the temporal setup, the number of horizons JJ (Sec. 3.3) follows the teacher clips of each backbone. On SimplerEnv (J=3J=3), an eight-frame teacher clip (consecutive offsets 0–7) accompanies a seven-action chunk. Two-frame tubelets give the three horizons as blocks (2,3), (4,5), and (6,7), with nominal tubelet midpoints at 0.83/1.50/2.17 s (Google) and 0.50/0.90/1.30 s (WidowX). All three lie within the seven-action prediction window, aligning the supervision horizons with the predicted action sequence. During execution, one temporally ensembled action is applied per observation. For π0.5 (J=2J=2), teacher pairs (Ii,Ii+6)(I_i,I_{i+6}) and (Ii,Ii+31)(I_i,I_{i+31}) define the two horizons, where ii indexes dataset timesteps. The short horizon lies within the ten-step prediction window but beyond the five-step execution prefix, while the long horizon extends beyond the prediction window. As actions after the chunk are not given to the predictor, this target is not determined by the chunk alone: it regresses the recorded demonstration future given the current scene, qtq_t, and the chunk, serving as a longer-range target under the demonstrated continuation. Evaluation uses 20 Hz control, with replanning after five actions. For training budgets, a single SimplerEnv model is trained jointly on Fractal (87,212 trajectories) and BridgeV2 (53,192) for 30,000 steps at global batch 256. LIBERO uses 1,693 no-op-filtered demonstrations for 30,000 steps at global batch 128, and LIBERO-Plus is evaluation-only. For the real-robot experiments, we use Qwen3-VL-2B-Instruct with a GR00T-style flow expert, trained on 100 demonstrations per task (600 in total) for 10,000 steps at global batch 256. Our models are trained on eight A800 GPUs with AdamW (β1,β2=0.9,0.95\beta_1,\beta_2=0.9,0.95, weight decay 0.01) and peak learning rates of 2×10−62\times10^{-6} (last eight language layers), 10−510^{-5} (action expert), 4×10−64\times10^{-6} (geometric predictor), and 1.5×10−61.5\times10^{-6} (residual adapter). The visual backbone and teacher stay frozen.

For the matched StarVLA baseline, "matched" means the same backbone, action expert, training data, batch size, number of updates, initialization, trainable layers, optimizer settings, and checkpoint-selection and evaluation protocols. It differs only in being trained with action supervision alone, without the predictor, adapter, or geometric objectives. For evaluation, each of our simulation results is the mean over three evaluation seeds of the same checkpoint, with 2,000 (LIBERO), 10,030 (LIBERO-Plus), and 864/96 (Google/WidowX) episodes per seed (App. C.4).

Q3: End-to-End Latency per Action Chunk.

Table R2 (Table 15 in our revised appendix). Latency per action chunk and SimplerEnv success (%).

PolicyLatency (ms)GoogleWidowX
Base policy (StarVLA)61.0057.655.2
GeoWRA67.5469.171.9

Both policies run on a single NVIDIA A100 GPU, with end-to-end latency including V-JEPA encoding for GeoWRA. GeoWRA adds only 6.54 ms (10.7%) for gains of 11.5/16.7 points on Google/WidowX. The extra time comes mainly from the world predictor and residual adapter, which run at each of the four flow steps shared by both policies, whereas V-JEPA encoding runs in parallel with Qwen encoding. In real-world deployment, we interpolate between consecutive predicted actions for smoother execution.

We hope these responses address your concerns and are glad to answer further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.