acceptodds
Under review as a conference paper at ICLR 2027

Evo-World: A Lightweight Vision-Language-Action Model with Future Motion Prediction via Optical-Flow Supervision

40 views

Abstract

Vision-language-action (VLA) models provide a unified framework for visual understanding, language grounding, and robot control, yet anticipating how the scene will evolve beyond the current observation can provide additional information for control. Action supervision alone does not explicitly encourage such anticipation, while future-image prediction entangles task-relevant motion with largely static appearance. To address these limitations, we introduce **Evo-World**, a lightweight VLA model that augments policy learning with future motion prediction through optical-flow supervision, thereby providing a task-relevant predictive signal focused on motion rather than appearance. We further design the training procedure to preserve these predictive representations during policy learning, regulating representation diversity while decoupling representation learning from action optimization. Across simulation benchmarks, Evo-World demonstrates strong performance, achieving a tier-averaged success rate of **91.1%** on Meta-World MT50 , with a 13.0% relative improvement over the non-predictive baseline. On real-world AgileX robots, Evo-World outperforms the evaluated baselines across single-arm and bimanual manipulation tasks while using only **0.87B** parameters and delivering lower GPU memory usage and higher inference efficiency. These results demonstrate that future motion prediction through optical-flow supervision provides an effective and lightweight approach to enhancing the predictive capabilities of VLA models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Vision-language-action models, Robotic manipulation, Motion prediction, Optical-flow supervision, Representation learning

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_FSS6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Evo-World is a lightweight vision-language-action model that learns predictive motion representations using optical flow supervision. The model contains a compact branch that predicts the embedding of future image-plane motion based on current observations and language instructions. The current embedding and the predicted embedding are used to modulate the policy through bounded FiLM modulation. The SIGReg mechanism is used to suppress representation collapse, while gradient detachment separates the motion representation learning from the action optimization process. At the inference stage, the model neither needs to perform optical flow estimation nor reconstruct future images. This paper evaluates the method on Meta-World MT50, RoboTwin, and real-world manipulation tasks, and conducts ablation experiments on the prediction targets, modulation features, fusion mechanisms, and optimization components.

Critical strengths

The results clearly show that adding motion information helps the policy. The model uses optical flow only to construct the training objective and makes predictions in the latent space at the deployment stage. Its total parameter count is 870 million (0.87B), an attractive design point: it gains motion prediction without running an optical flow estimator or reconstructing future observations during action generation.

Critical weaknesses

The related methods discussed in Section 2.3 are not compared with the method in this paper, so the empirical advantage of this approach over them has not been established. At minimum, controlled comparisons with the closest feasible methods, while reporting both task performance and computational cost, would make the empirical claims much more convincing.

The motivation for choosing optical flow is not very strong. Adding intermediate visual information in the middle of the VLA is a common practice. Comparing against other task-related approaches would give a stronger case for optical flow.

Questions for authors

Does "representation diversity" refer to preventing representation collapse and maintaining the differences between observations, or does it refer to semantic or motion-pattern diversity in a broader sense? The results of the SIGReg ablation experiment and Appendix C support the former.

The paper cites FiLM, but can it further explain the sources and rationale of the other design choices in the conditioning interface, especially the constrained 0.1tanh⁡0.1\tanh modulation and zero initialization? Which of these are inherited implementation choices, and which are the novel contributions of this paper?

3 comments
AuthorsOpenReview ↗

Response to Reviewer FSS6

We thank you very much for the constructive suggestions.

Response to Weaknesses

W1 — Missing comparisons with closely related methods and their computational costs: To our knowledge, Evo-World is the first work to introduce optical-flow supervision into a VLA with fewer than 1B parameters. Our study focuses on the effect of auxiliary motion supervision in lightweight models, complementing predictive VLA research at larger model scales. We therefore emphasize comparisons with lightweight baselines. In Table 1, the 0.77B-parameter Evo-1 base model achieves 80.6% on Meta-World, while the 0.87B-parameter Evo-World achieves 91.1%, a gain of 10.5 percentage points. Table 1 also includes similarly sized models such as Evo-Depth and FabriVLA.

Evo-World is trained on four consumer-grade RTX 4090 GPUs. At inference, it predicts a motion embedding from the current observation and instruction to condition action generation, without online optical-flow estimation or reconstruction. On an H20 with BF16 precision, it runs at 6.16 Hz with 2.54 GB GPU memory usage. The demonstrated training setup and low inference memory requirement support practical use on consumer GPUs, helping academic laboratories study embodied intelligence with accessible computing resources.

W2 — Insufficient justification for optical flow over other intermediate visual targets: Optical flow provides displacement-focused supervision, reducing the emphasis on static appearance in the prediction target. We have added sparse frame difference (FrameDiff) as another prediction target, alongside the original comparisons in Table 3A:

Prediction targetEasy (%)Medium (%)Hard (%)Very Hard (%)Tier-averaged SR (%)
None89.276.877.279.280.6
Future RGB90.780.078.388.084.3
FrameDiff92.979.176.778.081.7
Future optical flow (Evo-World)93.289.188.394.091.1

Optical flow improves success by 6.8 percentage points over future RGB and 9.4 points over FrameDiff. Frame differences indicate pixel-level changes, whereas optical flow explicitly describes displacement direction and magnitude. The additional comparison supports the value of this more structured motion target for action learning.

Response to Questions

Q1 — What does “representation diversity” mean? We mean preventing representation collapse and preserving differences across observations. We will revise the wording to make this definition explicit. SIGReg regularizes the current and target embeddings; the prediction loss connects the predicted embedding to the regularized target. Appendix C shows substantial representation contraction without SIGReg, and the corresponding success rate drops from 91.1% to 85.8%. These findings support its role in maintaining observation-dependent variation. In future work, we plan to add semantic probing experiments to assess how well the representations distinguish task semantics and motion categories, providing a more direct evaluation of their broader semantic and motion diversity.

Q2 — Why use bounded FiLM modulation and zero initialization, and which choices are inherited? The modulation is defined as (γ, β) = 0.1 × tanh(G([sg(z_t), sg(z_pred)])), followed by H' = H ⊙ (1 + γ) + β. The bound limits each channel's multiplicative factor to [0.9, 1.1] and additive shift to [-0.1, 0.1]. Zero-initializing the generator's final layer gives γ = β = 0, so conditioning is initially the identity transformation. These choices are intended to limit perturbations to pretrained features and allow the policy to learn modulation gradually; 0.1 is the modulation-scale hyperparameter.

FiLM follows Perez et al. (2018). We treat the bounded modulation and zero initialization as implementation choices, without claiming them as separate methodological contributions. Our contribution centers on learning and using predicted motion representations within a lightweight VLA.

Thank you for the clarification and for adding the FrameDiff experiments. This comparison largely alleviates my concern about the optical-flow choice; moreover, the explanation regarding representation diversity and the choice of FiLM implementation address my remaining doubts. I now have one main concern left: the comparison with closely related motion-supervised methods. Could you try selecting a representative existing method and evaluating it under the same lightweight backbone and a comparable training budget?

AuthorsOpenReview ↗

Response to Reviewer FSS6: Comparison with ATM

Thank you for the specific suggestion. We have added a comparison with ATM (Any-point Trajectory Modeling for Policy Learning), a representative open-source method that uses predicted point trajectories to guide policy learning, by adapting it to the same lightweight Evo-1 backbone.

Following ATM's trajectory-guided policy learning paradigm, we independently pretrain a trajectory predictor on Meta-World and keep it frozen during subsequent policy adaptation. Specifically, we initialize Evo-1 + ATM from an Evo-1 model trained for 100k steps, incorporate the ATM trajectory predictor, and perform an additional 10k steps of policy adaptation. This gives a cumulative policy-training budget of 110k steps, matching Evo-World’s budget in terms of policy optimization steps.

Our adaptation preserves ATM’s early- and late-conditioning design. Predicted trajectory tokens condition the action features before the action Transformer, while predicted trajectory coordinates additionally condition the final action MLP. These interfaces are adapted to Evo-1’s architecture, retaining its VLM backbone and flow-matching objective. The trajectory predictor remains frozen throughout policy adaptation.

We evaluate the adapted policy on all 50 Meta-World tasks, with 10 episodes per task, using the same evaluation protocol and tier-averaged success metric as in our paper.

The results are summarized below:

MethodEasy (%)Medium (%)Hard (%)Very Hard (%)Tier-averaged SR (%)
Evo-189.276.877.279.280.6
Evo-1 + ATM93.278.278.388.084.4
Evo-World93.289.188.394.091.1

The ATM-adapted baseline improves upon Evo-1 by 3.8 percentage points, confirming that predicted trajectory guidance can benefit the lightweight policy. Nevertheless, Evo-World’s score is numerically 6.7 percentage points above the ATM aggregate.

This additional comparison broadens our evaluation to a representative trajectory-guided alternative on the same lightweight backbone and provides supplementary empirical support for Evo-World’s effectiveness. We will incorporate this additional comparison with ATM into the final version of our paper.

Reviewer_N33ROpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents Evo-World, a lightweight vision-language-action policy augmented with a latent future-motion prediction branch. Given a current image and instruction, the branch predicts the DINOv2 embedding of a RAFT optical-flow rendering between frames (t) and (t+15). SIGReg is used to discourage representation collapse, action gradients are blocked from entering the motion branch, and current and predicted representations condition the VLA backbone through bounded FiLM modulation. The model reports 91.1% tier-averaged success on Meta-World MT50, 41.8% average success on RoboTwin, and 63.3% average success over three real-robot tasks. The main claimed benefit is that optical-flow supervision provides a task-relevant predictive signal while retaining low deployment cost.

Critical strengths
  1. The method is simple and deployment-oriented. It uses optical flow only as offline training supervision, while the deployed 0.87B model requires neither future observations nor online flow estimation.

  2. The paper evaluates several relevant design choices. Table 3 reports substantial configuration-level effects for the prediction target, FiLM conditioning, SIGReg, and gradient detachment, and the authors appropriately acknowledge limitations of the linear-probe evidence.

Critical weaknesses
  1. The paper does not demonstrate that the predicted representation adds future-motion information. Table A3 (p. 17) reports that the current representation z_t predicts future end-effector displacement better than the purported future representation z_pred: R^2=0.661 versus 0.508, and median angular error 20.9 degrees versus 29.4 degrees. Meanwhile, Table 3B compares concatenating z_t and z_pred against single-input conditions but provides no dimension-matched control such as z_t plus noise, a duplicate, or a random auxiliary branch. The gain can therefore be explained by additional capacity, auxiliary regularization, or encoding of expert action/task phase rather than anticipation. This is central to the paper's claimed mechanism, and no no-prediction/RGB/flow ablation is reported on RoboTwin or the real robot.

  2. The evidence for the proposed mechanism is confined to Meta-World. RoboTwin and the real-robot experiments evaluate only the complete Evo-World system; they do not compare optical-flow prediction with no prediction or future-RGB prediction. Consequently, these experiments show that the complete system performs well, but not that future-motion prediction causes the gains outside Meta-World. This substantially weakens the paper's claimed cross-domain contribution because the central mechanism is never isolated in the settings used to argue simulation generalization and real-world value.

Questions for authors
  1. The central claim is that predicting future motion provides useful information beyond the current observation. However, Table 3A removes the entire predictive objective and branch, Table 3B changes the conditioning content and dimensionality, and Table A3 explicitly states that it does not establish whether zpredz_{pred} adds information beyond ztz_t. Can the authors provide a matched intervention that keeps the optical-flow-supervised encoder, FiLM input dimension, architecture, and optimization fixed, but destroys only the future correspondence of zpredz_{pred}—for example by replacing it with duplicated/current features or a temporally mismatched prediction—and compare it with the valid future prediction? This is necessary to determine whether the reported gain is actually caused by future-motion prediction rather than auxiliary supervision or additional conditioning capacity.

  2. The predictor receives only the current observation and instruction, but its target is the optical flow produced by one demonstrated future trajectory. For the same observation and instruction, multiple valid action sequences can induce different future flows; an MSE predictor can therefore learn the dataset's dominant action or task phase rather than environment dynamics, or average incompatible futures. How does Evo-World behave when the demonstrated future is multimodal? Please evaluate whether zpredz_{pred} distinguishes different valid futures from the same state and whether its benefit persists on tasks requiring genuinely different action choices from visually similar states.

3 comments
AuthorsOpenReview ↗

Response to Reviewer N33R

We thank you for the careful analysis of the mechanism and its supporting evidence.

Response to Weaknesses

W1 — The evidence does not establish an incremental contribution from future-motion prediction: Table 3B ablates the FiLM conditions used during training and inference to examine the contribution of predicted motion representations to action generation. With optical-flow supervision retained, removing predicted features and conditioning only on current features z_t yields 83.7%, whereas the joint condition [z_t, z_pred] achieves 91.1%, an improvement of 7.4 percentage points. This supports the utility of predicted motion features as additional policy conditions. The new zero-padding and duplicated-current-feature controls in Q1 further control the FiLM input dimension to examine this benefit.

W2 — The mechanism has not been isolated beyond Meta-World: Given time constraints, we concentrated the comprehensive component and design ablations on Meta-World to analyze their contributions under a common experimental setup. RoboTwin and real-robot experiments focus on comparing the complete Evo-World system with its base model and other methods, evaluating overall performance in additional environments and tasks. In the expanded real-robot evaluation with 50 trials per task, Evo-1 and Evo-World achieve average success rates of 52.7% and 65.3%, respectively. We will clarify these experimental roles: Meta-World provides component and design analysis, while RoboTwin and real-robot experiments provide complete-system comparisons.

Response to Questions

Q1 — Can the contribution of correct future correspondence be tested while keeping supervision, architecture, and optimization unchanged? We added zero-padding and duplicated-current-feature controls that retain optical-flow supervision, replacing [z_t, z_pred] with [z_t, 0] and [z_t, z_t], respectively. All three conditions use a 2D-dimensional FiLM input, with consistent conditioning during training and evaluation.

Policy conditionEasy (280)Medium (110)Hard (60)Very Hard (50)Tier-Avg
[z_t, z_pred], Evo-World93.289.188.394.091.1
[z_t, z_t], duplicated current features92.9 (260/280)71.8 (79/110)61.7 (37/60)84.0 (42/50)77.6
[z_t, 0], flow supervision retained77.1 (216/280)69.1 (76/110)50.0 (30/60)80.0 (40/50)69.1

Entries report SR (%); parentheses give successes/trials for the added controls. Evo-World results are taken from the paper, including its reported Tier-Avg of 91.1%. Tier-Avg denotes the unweighted mean across the four buckets.

The full model outperforms both controls, improving on duplicated current features by 13.5 percentage points. Within these configurations, retaining auxiliary flow supervision or duplicating current features is insufficient to recover the full model's performance. These results support the practical contribution of predicted motion representations as action conditions.

Q2 — How does the model handle multiple valid futures for the same observation and instruction? Evo-World models future motion in a latent space, providing a common representation for different optical-flow patterns across observations and instructions. It learns flow embeddings and uses them as compact conditions for action generation without reconstructing full flow fields. This design aims to learn reusable motion structure from different demonstrations.

A latent space can represent different motion patterns, but this alone does not explicitly model multiple futures for identical inputs. With the same observation and instruction, the current deterministic predictor still outputs a single representation; with a fixed target representation, squared-error training may conditionally average different future modes. We therefore position the branch as an instruction-conditioned motion prior useful for control, without claiming that it explicitly distinguishes or samples multiple valid futures under identical inputs.

Thank you for the response and the additional controls. The new dimension-matched comparisons show that the improvement cannot be explained solely by increased FiLM input dimensionality or duplicated current features, and confirm that zpredz_{\mathrm{pred}} makes an additional contribution to policy learning. This resolves the central concern in my original review, and I have therefore changed my rating from 2 to 3.

AuthorsOpenReview ↗

Response to Reviewer N33R

Thank you for your thoughtful reassessment and constructive feedback. We’re glad the additional comparisons addressed your concern and helped strengthen our work.

Reviewer_N43uOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Evo-World is a 0.87B-parameter VLA built on Evo-1. It adds a branch that predicts the DINOv2 embedding of a future optical-flow rendering from the current image and instruction. SIGReg prevents representation collapse, and the predicted embedding conditions the policy through FiLM with gradients detached. It reports 91.1% on Meta-World MT50, 41.8% on RoboTwin, and 63.3% on three real-robot tasks (vs. 51.7% for π0), with lower memory use and faster inference.

Critical strengths

1.The method is simple, needs no flow estimation at inference, and is efficient to deploy (2.54 GB, 6.16 Hz).

  1. The real-robot comparison is controlled (same demonstrations, same GPU), and the ablations cover the main design choices.
Critical weaknesses

1.The gain from motion prediction is not isolated. Without SIGReg, the motion representations nearly collapse to constants (App. C), yet this variant still outperforms “None” by 5.2 points (Table 3). App. D also shows that z_pred encodes future motion worse than z_t. It is therefore unclear how much of the improvement comes from motion prediction itself.

2.The margins are small and the baselines are insufficient. No seeds or variances are reported, and the lead over FabriVLA on Meta-World is 1.1 points. The strongest real-robot baseline is π0; π0.5 is missing, and there is no Evo-1 baseline on the real robot. Novelty over FlowVLA and Fang et al. (2025), which also use optical flow as a prediction target, is limited, and the paper’s specific choices (a DINOv2 target, feeding z_pred to the policy at inference) are not ablated.

Questions for authors

1.Why does the no-SIGReg variant, whose representations nearly collapse, still outperform “None” by 5.2 points? Please report results over multiple seeds.

2.Can you add a variant that uses the flow loss during training but does not feed z_pred to the policy at inference?

3.Can you add π0.5 and Evo-1 to the real-robot comparison?

3 comments
AuthorsOpenReview ↗

Response to Reviewer N43u

We thank you very much for your careful evaluation.

Response to Weaknesses

W1 — The contribution of motion prediction is not sufficiently isolated: We agree that the no-SIGReg comparison and Table A3 do not fully isolate the contribution of predicted features. Table A3 evaluates each representation separately: current features achieve higher R² than predicted features, 0.661 versus 0.508. It does not test whether predictions improve decoding when current features and the instruction are already available. We will qualify this interpretation and use the completed matched control in Q2 to further assess the role of predicted features as policy conditions.

W2 — Statistical evidence, baselines, and novelty: All reported Evo-World simulation evaluations are averages over multiple seeds. We have also added Evo-1 and π₀.₅ and expanded real-robot evaluation to 50 trials per task (Q3).

Our work focuses on integrating optical-flow-supervised motion prediction into a lightweight VLA with only 0.87B parameters. Evo-World directly predicts motion embeddings from the current observation and instruction and uses them to condition action generation through bounded FiLM, requiring no future observations, online optical-flow estimation, or flow reconstruction. Predicted motion embeddings remain available as policy conditions at inference, while motion learning is separated from action optimization: the prediction loss and SIGReg train the motion branch, and action gradients update only the conditioning module and policy. FiLM and SIGReg are established components; our contribution is how motion representations are learned and used within a lightweight VLA.

Response to Questions

Q1 — Why does no-SIGReg outperform None by 5.2 points, and is this consistent across seeds? Our interpretation is that this gain may reflect the optimization history. Even if representations eventually collapse, non-collapsed motion representations early in training may influence policy optimization through FiLM conditioning, leading to a different training trajectory from the None baseline. Representation contraction at the final checkpoint therefore does not rule out an earlier contribution from the motion branch. This is a possible explanation, rather than a mechanism directly established by training-dynamics diagnostics.

The table reports mean success rates and standard deviations over five seeds. Without SIGReg, the mean remains 5.2 points above None, while the full model further improves to 91.1%.

ConfigurationTier-averaged SR (%)SeedsSD (percentage points)
None80.650.9
Without SIGReg85.851.2
Evo-World91.150.7

Q2 — Can flow supervision be retained without using predicted features at inference? Yes. We added zero-padding [z_t, 0] and duplicated-current-feature [z_t, z_t] controls, retaining the optical-flow loss, SIGReg, and gradient separation. All three conditions use a 2D-dimensional FiLM input. Each control uses the same replacement during training and evaluation, preserving auxiliary flow learning while removing predicted-feature conditioning and avoiding test-only input changes.

Policy conditionEasy (280)Medium (110)Hard (60)Very Hard (50)Tier-Avg
[z_t, z_pred], Evo-World93.289.188.394.091.1
[z_t, z_t], duplicated current features92.9 (260/280)71.8 (79/110)61.7 (37/60)84.0 (42/50)77.6
[z_t, 0], flow supervision retained77.1 (216/280)69.1 (76/110)50.0 (30/60)80.0 (40/50)69.1

Entries report SR (%), with successes/trials in parentheses for the controls. Evo-World uses the paper's results; Tier-Avg denotes the equal-weight bucket average.

Evo-World achieves 91.1%, exceeding zero-padding (69.1%) and duplicated current features (77.6%) by 22.0 and 13.5 percentage points, respectively. In these configurations, auxiliary flow supervision alone or duplicated current features does not recover the full model's performance, supporting the utility of predicted motion features as policy conditions.

Q3 — Can Evo-1 and π₀.₅ be added to the real-robot comparison? Yes. Both models have been added, with 50 trials per task. The average weights all three tasks equally.

MethodFloral arrangement (%)Mug hanging (%)Microwave placement (%)Average SR (%)
OpenVLA-OFT22.024.018.021.3
π₀38.044.056.046.0
π₀.₅76.068.062.068.7
Evo-148.050.060.052.7
Evo-World58.074.064.065.3

Although π₀.₅ achieves a higher average success rate across tasks than Evo-World, it has substantially more parameters (3.3B vs.0.87B). Evo-World achieves the highest average success rate among all models with fewer than 1B parameters evaluated in our experiments.

AuthorsOpenReview ↗

Kind Follow-up on Our Responses Before Discussion Closes

Dear Reviewer N43u,

As the discussion period is drawing to a close, we would greatly appreciate any further feedback you may have on our responses. We would be happy to provide additional clarification and address any remaining questions or concerns. If our responses have adequately addressed your concerns, we would be grateful if you could consider raising your score. Thank you for your time and consideration.

Thank you for the additional experiments. The dimension-matched controls address my concern that the gains might come simply from a larger conditioning input or from repeating the current features. The comparison with ATM on the same backbone also strengthens the case for the proposed method in the evaluated setting.

My remaining concern is how broadly that advantage holds over existing motion-supervision methods. The ATM comparison is limited to one setting, and the real-robot results show an improvement over Evo-1 alone (65.3% vs. 52.7%) but no improvement over π0.5 (68.7%). Taken together, the new results support the usefulness of the predictive features, but they do not yet establish a broader advantage for this particular approach. I therefore maintain my original rating.

Reviewer_Uj6eOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Evo-World is a VLA policy with a future-motion prediction branch. During training, the VLA policy additionally predicts the RAFT optical flow embedding for future motion supervision.

At test time, the branch predicts this embedding from the current image and instruction. Current and predicted features condition the action policy through FiLM. SIGReg discourages representation collapse, while a stop-gradient blocks action gradients from entering the motion branch. The paper reports 91.1% tier-averaged success on Meta-World MT50, 41.8% averaged over RoboTwin Clean and Random, and 63.3% over three real robot tasks.

The model is reported to use 0.87B parameters. The ablations compare prediction targets, conditioning inputs, fusion methods, and training components.

Critical strengths
  1. The prediction target is well motivated and supported by a controlled comparison. Optical flow focuses the target on motion. Table 3A reports 91.1% for flow supervision, versus 84.3% for future RGB and 80.6% without prediction (p. 8). The policy does not need future observations or a flow estimator at test time (Sec. 3.3, pp. 4–5).
  2. The method is deploy-friendly. While maintaining competitive performance, the proposed method has less than 1B parameters and could be deployed on consumer GPUs over robot platform rather than cloud server.
Critical weaknesses
  1. The evidence does not show how stable the reported gains are. Each real task has only 20 trials, and no uncertainty estimates or repeated training runs are reported. The 65% vs 60% result on microwave placement differs by just one success. The larger simulation comparisons and ablations also lack variation across runs.
  2. The simulation results are weak. The Robotwin Clean2Random result has an obvious gap between the SOTA methods.
  3. The novelty of predicting visual latent is insufficient. There are many previous works that add future visual latent prediction into the VLA backbone. Directly change DINO/Megaflow feature to RAFT features lacks novelty.
Questions for authors
  1. The authors should claim the core novelty and difference with previous works.
  2. Which Table 1 results were rerun, and which were quoted? For reruns, were data, inputs, training budgets, and evaluation scripts matched? Please explain the remaining differences.
  3. Why does concatenation fall from 81.4% with future RGB targets to 53.2% with flow targets, while FiLM improves? Were projection sizes, parameter counts, initialization, and optimization matched?
2 comments
AuthorsOpenReview ↗

Response to Reviewer Uj6e

We thank you very much for your careful evaluation.

Response to Weaknesses

W1 — Stability of the reported gains: All reported Evo-World simulation evaluations average multiple seeds. Since 20 real-robot trials per task provide limited precision, we expanded evaluation to 50 trials per task and added Evo-1 and π₀.₅. All success rates below are percentages; the average weights the three tasks equally.

MethodFloral arrangementMug hangingMicrowave placementAverage SR
OpenVLA-OFT22241821.3
π₀38445646.0
π₀.₅76686268.7
Evo-148506052.7
Evo-World58746465.3

π₀.₅ achieves a higher average success rate but has substantially more parameters (3.3B vs. 0.87B). Evo-World has the highest average success rate among the evaluated models below 1B parameters.

W2 — Limited Clean2Random performance: Evo-World achieves 16.5% on RoboTwin Clean2Random, below the best result of 18.2% in Table 1, but leads among the evaluated models below 1B parameters. Our goal is to enable training and inference on consumer GPUs available in research laboratories, lowering the computing barrier to embodied AI research and development.

W3 — Novelty beyond future visual-latent prediction: Future visual-latent prediction and optical-flow supervision have established precedents. Our contribution is bringing optical-flow-supervised predictive representation learning to lightweight VLA models. To our knowledge, Evo-World is the first lightweight VLA to learn predictive motion representations through optical-flow supervision for action generation.

RAFT computes flow targets offline from demonstration frames, while a DINOv2-initialized encoder encodes current images and flow renderings. The predictor regresses the learned embedding of a flow rendering, shifting the target from future appearance to image-space displacement. RAFT supplies supervision; DINOv2 remains the representation backbone. Q1 explains how these representations inform the policy.

Response to Questions

Q1 — What is the core contribution and difference from prior work? Evo-World integrates optical-flow-supervised motion prediction into a 0.87B VLA. It predicts motion embeddings from the current observation and instruction and uses them to condition action generation through bounded FiLM. These embeddings remain policy conditions at inference, requiring no future observations, online flow estimation, or flow reconstruction.

Motion learning is separated from action optimization: the prediction loss and SIGReg train the motion branch, while action gradients update only the conditioning module and policy. FiLM and SIGReg are established components; our contribution is how motion representations are learned and used within a lightweight VLA.

Table 3A supports the target choice: success rates are 80.6% without prediction, 84.3% with future RGB, and 91.1% with optical flow.

Q2 — Which Table 1 results were rerun or quoted, and were conditions matched? We evaluated Evo-1 and Evo-World. All other Table 1 results come from the respective models' official papers and were not rerun under our training pipeline. For each benchmark, the evaluations use the same simulation environment implementation from the official repository.

Q3 — Why does concatenation degrade with flow targets while FiLM improves, and were configurations matched? We report an MT50 macro-average across four difficulty tiers, each weighted at 25%, rather than a micro-average over all 500 episodes. Very Hard contains 10% of episodes but receives 25% of the weight (2.5×); Hard contains 12% but receives 25% (about 2.1×); Easy contains 56% but receives 25% (about 0.45×). Thus, concat's weaker results on multi-step tasks—34.5% on Medium and 40.0% on Very Hard—have greater influence, yielding 52.3% macro-average versus 64.4% micro-average.

Directly concatenating DINOv2 flow features with VLM features reduces success on difficult tasks. A possible explanation is representation mismatch: DINOv2 was not explicitly pretrained on flow renderings, so its flow embeddings may align poorly with VLM features. Direct concatenation may consequently interfere with action-relevant representations, especially on challenging tasks.

AuthorsOpenReview ↗

Kind Follow-up on Our Responses Before Discussion Closes

Dear Reviewer Uj6e,

We sincerely appreciate the time and care you have devoted to reviewing our submission. As the discussion period closes tomorrow, we would be grateful if you could take a moment to review our responses, including the expanded real-robot evaluations and clarifications of our contributions. If any concerns remain, we would be happy to clarify further or provide additional analysis within the remaining time. If our responses have sufficiently addressed your concerns, we would greatly appreciate your considering a higher rating. Thank you again for your thoughtful feedback.

Comments on the paper

2 comments
Area_Chair_Rk4AOpenReview ↗

Please Review the Author Rebuttal and Reassess Your Rating

Dear Reviewers,

Thank you for your time and effort in reviewing the submissions.

The authors have now submitted their rebuttals. I kindly ask you to carefully review the authors' responses and assess whether they have adequately addressed your questions and concerns.

If the rebuttal resolves your concerns, please consider updating your review and adjusting your rating accordingly. If any concerns remain, please feel free to raise them in the discussion.

Your timely feedback and active participation are greatly appreciated, as they are essential to reaching a fair, well-informed decision.

Thank you again for your valuable contributions!

Best regards, Area Chair

AuthorsOpenReview ↗

Rebuttal Summary

Dear Area Chair and Reviewers,

We thank the Area Chair and all four reviewers for their time and thoughtful feedback. During the discussion, three reviewers provided follow-up comments on our rebuttal, with encouraging and positive feedback on the additional experiments and clarifications:

  • Reviewer N33R confirmed that the new dimension-matched controls resolved the central concern and raised the rating from 2 to 3.
  • Reviewer N43u acknowledged that the dimension-matched controls addressed the concerns about increased conditioning dimensionality and duplicated current features, and maintained their original positive recommendation (3).
  • Reviewer FSS6 acknowledged that FrameDiff and the design clarifications addressed several concerns and requested a related-method comparison; we provided an ATM adaptation and are awaiting feedback on that addition.

Across the reviews, reviewers highlighted several strengths of Evo-World:

  • Lightweight and efficient deployment: the 0.87B-parameter model offers a practical deployment profile, with reported inference memory usage of 2.54 GB and a rate of 6.16 Hz.
  • No online optical-flow estimation: optical flow is used as offline training supervision, while action generation does not require future observations, online flow estimation, or future-image reconstruction.
  • Motion-focused supervision: reviewers noted evidence from the prediction-target comparison and ablations that motion information can benefit policy learning.
  • Coverage of key design choices: the ablations examine prediction targets, conditioning features, fusion mechanisms, and optimization components.

The additional experiments and analyses are summarized below:

  • Dimension-matched controls: We added controls using duplicated current features and zero padding while retaining optical-flow supervision and FiLM input dimensionality, supporting the contribution of predicted motion features to policy learning.
  • Prediction-target comparison: We added FrameDiff alongside future RGB and optical-flow targets, providing additional evidence for the value of displacement-focused supervision.
  • Related-method comparison: We adapted ATM to Evo-1 using a pretrained, frozen trajectory predictor and early/late conditioning, adding a comparison with an established trajectory-guided method on the same lightweight backbone.
  • Statistical reporting: We reported success-rate means and standard deviations for the baseline without prediction, the variant without SIGReg, and the full model.
  • Expanded real-robot evaluation: We expanded real-robot trials and added Evo-1 and π₀.₅ as baselines, showing improvements over Evo-1 while noting the remaining gap to the larger π₀.₅ model.

We appreciate the reviewers’ suggestions and will incorporate the additional experiments and clarified scope into the revised manuscript. Thank you again for your time!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.