acceptodds
Under review as a conference paper at ICLR 2027

OneVLA: A Unified Framework for Embodied Tasks

Abstract

Navigation and manipulation are fundamental capabilities of embodied intelligence, enabling robots to interpret natural language commands and interact physically with their surroundings. However, current Vision-Language-Action (VLA) models remain constrained by task-specific architectures, specializing in either navigation or manipulation, which hinders the development of general-purpose robotic agents. To bridge this gap, we introduce OneVLA, a unified architecture that integrates these distinct tasks into a single, cohesive framework. Specifically, we design a unified action head capable of generating both navigation and manipulation actions without requiring task-specific variants. Furthermore, we propose a multi-stage progressive training strategy—incorporating curated data construction and Chain-of-Thought (CoT) fine-tuning—that facilitates strong positive transfer and mutual reinforcement between the two domains. Extensive experiments in both simulated and real-world environments demonstrate that OneVLA achieves state-of-the-art performance, significantly outperforming both specialized single-task and existing cross-task models. By unifying these core capabilities, OneVLA paves the way for truly general-purpose robotic systems. The source code will be publicly released.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied Navigation, Embodied Manipulation, VLA

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_6Yr2OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces OneVLA, which unifies navigation and manipulation within a single framework. It proposes a unified action output head and a multi-stage training strategy intended to promote mutual improvement between these two tasks. The paper evaluates the framework on navigation and manipulation benchmarks, including mobile manipulation, and presents real-world experiments and ablation studies.

Critical strengths
  1. Unifying navigation and manipulation is a meaningful research direction

    OneVLA addresses both tasks within a shared architecture, and the reported improvements from joint training suggest that learning these capabilities together can be beneficial.

  2. The evaluation covers both simulation and real-world settings

    The additional mobile manipulation experiments on RoboCasa365 are particularly relevant, as they extend the validation beyond evaluating navigation and manipulation separately.

Critical weaknesses
  1. The baseline comparisons are not sufficiently strong for broad state-of-the-art claim

    For example, OpenVLA-OFT is included in the main comparison but omitted from the LIBERO results, where the paper instead emphasizes improvements over the original OpenVLA. Including stronger baselines under comparable settings would help establish whether the reported gains remain substantial.

  2. The contribution of the unified action head is not clearly isolated

    Although the paper includes joint-training ablations and a UniVLA variant with the proposed head, these do not directly compare a unified head against separate navigation and manipulation heads using the same backbone and training recipe. This makes it difficult to determine whether the improvements come from the proposed architectural design or from joint training and the multi-stage training strategy.

Questions for authors
  1. Baseline clarification

    Why is OpenVLA-OFT included in the main comparison but omitted from the LIBERO evaluation? Are there differences in evaluation settings that prevent a fair comparison, and does OneVLA retain its advantage under comparable settings?

  2. Evidence for the architectural contribution

    Have you compared the unified action head with separate navigation and manipulation heads using the same backbone, training data, and training strategy? Such results would help determine whether the unified head itself contributes to the improvements beyond joint training.

  3. Navigation evaluation clarification

    The paper reports OSR as the main navigation metric. Does OneVLA also outperform the baselines in SR and SPL, which account for successful stopping and path efficiency?

6 comments
AuthorsOpenReview ↗

Thanks and Response to Reviewer 6Yr2 (1/2)

We sincerely thank the reviewer for the thoughtful and constructive feedback. We greatly appreciate your recognition of our “meaningful research direction” and “particularly relevant RoboCasa365 experiments”. Below, we address your main questions and concerns in detail.


Q1: Query about comparison with OpenVLA-OFT.

Thanks for your question. We included OpenVLA-OFT in SimplerEnv. The original LIBERO table focused on a cross-task comparison with UniVLA. Table 14 adds the published LIBERO results for OpenVLA-OFT, π0, and π0-FAST.

Rebuttal Table 14. Expanded LIBERO baseline comparison. OneVLA's scores reproduce original Table 5. External scores come from the papers linked below.

MethodParamsModels used for the 4 suitesAlso does navigationSpatial ↑Object ↑Goal ↑Long ↑Avg. ↑
OpenVLA7B4 (one per suite)✗84.788.479.253.776.5
π0-FAST3B1✗96.496.888.660.285.5
π03B✗96.898.895.885.294.2
UniVLA7B4 (one per suite)✗96.596.895.692.095.2
OpenVLA-OFT7B4 (one per suite)✗97.698.497.994.597.1
OneVLA (Ours)3B1✓98.297.898.196.597.7

Evaluation settings. OneVLA achieved 97.7% average success and 96.5% on LIBERO-Long. OpenVLA-OFT reports 97.1% and 94.5%, respectively, with a separate 7B policy for each suite using third-person and wrist images plus proprioceptive state. OneVLA uses one 3B policy that also supports navigation. OpenVLA-OFT also reports 96.8% with a single policy across all four suites in its Table XIV, 0.9 points below OneVLA. We will include the stronger baselines and clarify their inputs and policy settings in the revision.


Q2: Query about the effectiveness of unified action head.

Thanks for your question. We compared a shared DiT action head with separate navigation and manipulation heads in both the single-stage and full-training settings. In each comparison we kept the backbone, data, and training recipe fixed.

(A) Same backbone and single-stage training. We used the setup from original Table 2b and added two controls, one with two full DiT-B heads and one with two heads whose combined parameter count matches the unified head. All joint models used Qwen2.5-VL-3B, the same data, a 6:4 task sampling ratio, and the same number of training steps. Table 15 shows the results.

Rebuttal Table 15. Separate-head controls extending original Table 2b. Single-task and unified-head scores are existing results. Both separate-head rows are new.

Action headHead paramsR2R OSR ↑RxR OSR ↑SimplerEnv ↑
Single-task training (Navi. only / Mani. only, original Table 2b)1×48.833.640.7
Shared backbone + separate navigation / manipulation heads2×53.439.643.5
Shared backbone + separate heads (param-matched)1×52.738.542.9
Unified head (Ours, original Table 2b)1×56.142.946.2

Joint training with separate heads shows how much the tasks gain from sharing the backbone. Compared with two full heads, the unified head adds 2.7, 3.3, and 2.7 points on R2R, RxR, and SimplerEnv. The parameter-matched comparison checks whether the gains hold at the same total head size.

(B) Same full training recipe. We also trained the separate-head and unified-head models on the same three-stage, 60K-step schedule. Table 16 shows the results.

Rebuttal Table 16. Head comparison under the complete training strategy. OneVLA's scores are from the original paper. The separate-head row is a new control.

Action headR2R OSR ↑RxR OSR ↑SimplerEnv ↑
Shared backbone + separate navigation / manipulation heads65.054.360.9
Unified head (OneVLA)68.658.264.5

The unified head improves R2R, RxR, and SimplerEnv by 3.6, 3.9, and 3.6 points, respectively. Original Table 8 reports a complementary experiment in which we transferred the unified head to UniVLA. We will add both controlled comparisons to the revision and explain what they show about the unified head.

AuthorsOpenReview ↗

Thanks and Response to Reviewer 6Yr2 (2/2)

Q3: Query about navigation metrics on VLN-CE.

Thanks for your question. OneVLA achieves higher SR and SPL than the listed baselines on both R2R-CE and RxR-CE.

(A) SR and SPL comparison. We measured NE, SR, and SPL on the R2R-CE and RxR-CE val-unseen splits and nDTW on RxR-CE. Tables 17 and 18 show these results alongside the original OSR and the baseline scores. The sources below each table identify the published results and the StreamVLN version.

Rebuttal Table 17. Expanded R2R-CE metrics. OneVLA's OSR is from original Table 1.

MethodParamsUnifiedNE ↓OSR ↑SR ↑SPL ↑
NaVILA7B✗5.2262.554.049.0
StreamVLN [ICRA’26]7B✗4.9864.256.951.9
UniVLA-Navi.* (OneVLA baseline)7B✗6.7147.138.234.5
OneVLA (Ours)3B✓4.7268.658.453.1

Rebuttal Table 18. Expanded RxR-CE metrics. OSR scores are from original Table 1.

MethodParamsUnifiedNE ↓OSR ↑SR ↑SPL ↑nDTW ↑
NaVILA7B✗6.7753.649.344.058.8
StreamVLN [ICRA’26]7B✗6.2255.752.946.061.9
UniVLA-Navi.* (OneVLA baseline)7B✗8.3526.321.418.237.9
OneVLA (Ours)3B✓5.9858.253.746.862.6

OneVLA leads StreamVLN v1 by 1.5 points in SR and 1.2 in SPL on R2R, and by 0.8 points on both metrics on RxR.

(B) How stopping is learned. We give the rare STOP command a loss weight of 15 and test this choice in original Table 7. The added SR and SPL results show how often the model stops successfully and how efficiently it completes each route. We will add these metrics and clarify the STOP weighting in the revision.


We hope these responses address your concerns, and we would be glad to answer any further questions.

AuthorsOpenReview ↗

To Reviewer 6Yr2

Dear Reviewer 6Yr2,

We hope this message finds you well. With the deadline of discussion period is approaching, we wanted to make sure we have fully addressed your concerns. If you have any further points or feedback you would like us to consider, please let us know. Your insights are invaluable to us, and we are open to any further suggestions you might have to improve the paper.

Thank you for the time and effort you have dedicated to reviewing our paper.

Best regards,

The Authors

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

This paper proposes OneVLA, a unified VLA framework for embodied tasks. Specifically, the authors build on Qwen2.5-VL-3B and introduce a unified 11-D action space that concatenates 4-D discrete navigation commands with 7-D end-effector deltas. A DiT flow-matching head generates actions in this space, with task-specific per-dimension loss masks. The model produces text reasoning before actions via special tokens and is trained with a three-stage progressive strategy: manipulation + VQA, then adding navigation and CoT data. Experimental results on VLN-CE R2R/RxR, SimplerEnv-Bridge, LIBERO, RoboCasa365 and real-robot tasks validate the effectiveness of the method.

Critical strengths
  1. Jointly training navigation and manipulation in a single model is an interesting idea.
  2. The method shows clear performance gains over baselines.
Critical weaknesses

1.No design that addresses the intrinsic differences between navigation and manipulation. The two tasks differ fundamentally. Navigation involves large egocentric viewpoint changes over long horizons and requires long-term memory, but only coarse decisions. Manipulation operates from a relatively fixed viewpoint over short horizons, but requires fine-grained, precise interaction actions. OneVLA unifies them mainly by zero-padding and concatenating action dimensions with per-dimension loss masks (Eq. 4–5). The same backbone, input resolution and action horizon are shared by both tasks. Navigation memory is handled only by uniformly sampling 7 history frames (App. A.1/B), and discrete navigation commands are modeled by flow-matching regression on one-hot vectors followed by argmax. The authors themselves note that the tasks have "different temporal scales" (App. F), yet no corresponding design is proposed. 2.No coupled mobile manipulation evaluation, and unclear why navigation improves manipulation. Navigation and manipulation are evaluated separately, and the real-world experiments even use different platforms (a mobile base vs. a fixed Franka arm). There is no instruction-following setting in which the robot navigates to different locations in a house and then manipulates with a fixed base. The only mobile manipulation results (RoboCasa365, App. C) are simulation-only, involve short-range base motion within one kitchen.

Questions for authors
  1. Have you considered task-specific designs, such as long-term memory for navigation or separate action horizons and resolutions for the two tasks? Why model discrete navigation commands with flow matching?
  2. Can you evaluate an instruction-following setting that requires navigating between rooms and then manipulating? How are RoboCasa365 base actions mapped to the unified action space?
4 comments
AuthorsOpenReview ↗

Thanks and Response to Reviewer knaj (1/2)

We sincerely thank the reviewer for the thoughtful and constructive feedback. We greatly appreciate your recognition of our “interesting idea” and “clear performance gains”. Below, we address your main questions and concerns in detail.


Q1: Query about task-specific memory, action horizons, and image resolution.

(A) Observations and execution. We use a shared backbone and action head with different inputs and action decoding for the two tasks. For navigation, we use the current frame and seven history frames, including the first frame. We sample the other frames along the preceding trajectory to track viewpoint changes and instruction progress. For manipulation, we use the current observation for local control. We decode navigation commands with argmax and apply denormalization and temporal action ensembling for manipulation. Section 3.2 describes how we weight inactive dimensions and the rare STOP command.

(B) Action horizon and training order. We tested horizons from two to eight steps. Five steps gave the best results on both R2R and SimplerEnv (original Figure 4). We also tested the order of training the two tasks. Training on manipulation first raised the Stage 2 mean by 11.5 points over training on navigation first (original Table 10), supporting our choice to learn fine-grained control before adding navigation.

(C) Comparison with task-specific alternatives. We changed one setting at a time by varying the number of navigation history frames, the action horizon, or the navigation resolution. We kept the rest of the training setup fixed and evaluated both tasks. Table 8 shows the results.

Rebuttal Table 8. Additional navigation design ablations. The default configuration and its OSR/SimplerEnv scores are from the original paper. Alternative configurations and SR entries are new.

VariantR2R OSR ↑R2R SR ↑RxR OSR ↑RxR SR ↑SimplerEnv ↑
OneVLA (default: 7 history frames, shared horizon 5, 224 px)68.658.458.253.764.5
Longer navigation memory: 15 history frames68.157.957.753.264.3
Longer navigation memory: 31 history frames (4× visual tokens)67.757.457.252.864.2
Separate horizons: navigation 3, manipulation 567.257.156.952.564.2
Separate horizons: navigation 8, manipulation 566.155.855.751.264.4
Higher navigation resolution (336 px)68.358.157.953.464.3

The default of seven history frames, a shared five-step horizon, and 224 px resolution gave the highest scores on all reported metrics across R2R, RxR, and SimplerEnv. We will describe these choices and include the ablations in the revision.


Q2: Query about flow matching head for navigation.

We use flow matching so that one action head can generate both discrete navigation commands and continuous manipulation actions. We compared it with a separate navigation classifier and with autoregressive text action tokens. All three setups used the same backbone, data, training schedule, and five-step action horizon. Table 9 shows the results.

Rebuttal Table 9. Navigation output formulations. The unified model's OSR and SimplerEnv scores are from the original paper. Alternative formulations and SR entries are new.

Navigation output# action headsR2R OSR ↑R2R SR ↑RxR OSR ↑RxR SR ↑SimplerEnv ↑
Cross-entropy classification head (per-step softmax)265.855.555.050.861.6
Text action tokens (autoregressive)1 + LM64.153.853.549.162.4
Flow matching in the unified space (Ours)168.658.458.253.764.5

Unified flow matching performed best on every reported metric. We will add this comparison to the revision.


We hope these responses address your concerns, and we would be glad to answer any further questions.

AuthorsOpenReview ↗

Thanks and Response to Reviewer knaj (2/2)

Q3: Query about mobile manipulation experimental results.

In our original RoboCasa365 evaluation (Appendix C, Table 6), the same policy moved the base and controlled the arm within one episode. OneVLA achieved 28.4% task-weighted success, compared with 20.0% for GR00T N1.5 and 9.9% for UniVLA. On Composite-Unseen, the scores were 7.9%, 4.4%, and 1.1%, respectively.

(A) Longer-range navigation before manipulation. We reran the same 50 RoboCasa365 tasks with the robot starting farther from the target. Table 10 compares success rates at the original and longer starting distances.

Rebuttal Table 10. Longer-range RoboCasa365 control. Standard-start scores reproduce original Table 6. Long-range scores are from the new setting.

InitializationUniVLAOneVLA
Standard (original Table 6)9.928.4
Long-range start5.224.6

Starting farther away reduced success by 13.4% relative to the standard start for OneVLA and by 47.5% for UniVLA. We will add the evaluation and the starting distances to Appendix C.

(B) Cross-room instruction following. In the real-world evaluation, the robot received the full instruction, navigated between rooms, and then manipulated with its base fixed. We tested four instructions with ten trials each for OneVLA, MapNav with Diffusion Policy, and oracle-switched UniVLA. Table 11 shows the results.

Rebuttal Table 11. Cross-room evaluation protocol. The evaluation uses 40 trials per method.

MethodTask switchingNavigation SR ↑Manipulation SR (given nav. success) ↑Overall SR ↑
MapNav + Diffusion Policy (modular)Hand-coded65.0 (26/40)61.5 (16/26)40.0 (16/40)
UniVLA (task-specific heads)Oracle switch45.0 (18/40)66.7 (12/18)30.0 (12/40)
OneVLA (single policy)None77.5 (31/40)80.6 (25/31)62.5 (25/40)

OneVLA navigated successfully in 31 trials and completed 25 instructions, for 62.5% overall success. This was 22.5 points above the modular baseline. The per-stage rates show where each method failed. We will add the evaluation, robot platform details, and per-stage results to Appendix C.

(C) Base and arm control. We use an adapter to map OneVLA's shared action representation to RoboCasa365's 12-dimensional control vector. Seven dimensions control the end effector and gripper (3D translation, 3D rotation, and gripper control). Three control base motion (planar translation and yaw), one controls torso height, and one indicates the control mode. We keep the loss weighting from Appendix C. During base movement, the manipulation dimensions receive a weight of 0.1. During arm manipulation, the navigation dimensions receive a weight of 0.1. We will document the adapter and the loss weighting in Appendix C.


Q4: Query about cross-task positive transfer and mutual reinforcement.

With the model and training steps fixed, adding navigation data raised SimplerEnv success from 40.7% to 46.2% (original Table 2b). We tested whether the gain comes from skills that both tasks use, such as tracking landmarks across changing viewpoints and interpreting spatial relations like “in” and “next to.”

(A) What navigation data contributes. We replaced the navigation data with an equal amount of general VQA or spatial VQA and kept the original single-stage training setup. Table 12 shows the results.

Rebuttal Table 12. Data-source controls extending original Table 2b. Manipulation-only and navigation-joint SimplerEnv scores are already reported. VQA controls and LIBERO scores are new.

Data added to manipulation trainingSimplerEnv ↑LIBERO Avg. ↑
None (Mani. only)40.792.5
+ General VQA (size-matched)41.492.8
+ Spatial VQA (size-matched)42.893.4
+ Navigation data (joint training, Ours)46.294.6

General VQA improved SimplerEnv by 0.7 points and spatial VQA by 2.1 points. Navigation data improved it by 5.5 points.

(B) Which manipulation tasks benefit. We also compared manipulation-only training and joint training on each LIBERO suite. We expected the largest gains on LIBERO-Spatial and LIBERO-Long if spatial understanding and sequential decision-making help explain the transfer. Table 13 shows the results.

Rebuttal Table 13. LIBERO transfer by suite under single-stage training.

Training settingSpatialObjectGoalLongAvg.
Mani. only93.295.693.487.892.5
Joint training96.096.295.091.294.6
Gain+2.8+0.6+1.6+3.4+2.1

The largest gains were on Long (3.4 points) and Spatial (2.8 points), consistent with that explanation. Both experiments use the single-stage setup in original Table 2b. The 97.7% LIBERO result in original Table 5 uses the full training pipeline. We will add both comparisons and discuss the transfer in the revision.

what's the Cross-entropy classification head (per-step softmax) for SimplerEnv? Could u provide the details for reimplementing the results of SimplerEnv-Bridge?

AuthorsOpenReview ↗

Thanks and Response to Reviewer KPaZ

Thanks for your question.


Q1: Query about Cross-entropy classification head and SimplerEnv reproduction.

The CE-head variant keeps OneVLA's backbone, data and training schedule. The only change is to the 4 navigation dimensions: instead of using flow matching, a linear layer predicts per-step logits over {forward, left, right, stop} for the 5-step action chunk, trained with softmax cross-entropy (stop weight 15). Manipulation still uses the flow-matching head, which is why this variant also reports a SimplerEnv score.

To reproduce our SimplerEnv-Bridge results:

  • Benchmark: the official WidowX visual-matching setting, 4 tasks × 24 episodes, max 【60/120】 steps.
  • Training data: BridgeData V2 under the oxe_bridge config with q01/q99 normalization. No SimplerEnv data is used.
  • Inputs: a single 224×224 third-person image and the instruction.
  • Action generation: a 5-step action chunk from 4 Euler denoising steps, with noise sampled from Beta(1.5, 1.0).
  • Post-processing: we un-normalize the actions with the q01/q99 statistics, apply exponential temporal ensembling (α = 0.1, K = 【5】), and binarize the gripper at 0.5.
  • Checkpoint: the same 60K-step checkpoint is used for every benchmark.

We will release the evaluation code and checkpoint.


We hope these responses address your concerns and would be glad to answer any further questions.

Reviewer_KPaZOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents OneVLA, a unified vision-language-action model for navigation and manipulation. The method concatenates four-dimensional discrete navigation action representations and seven-dimensional continuous manipulation actions into a unified action space, using a shared flow-matching action head and task-specific loss weights. Training proceeds in three stages: manipulation pretraining, cross-task joint training, and CoT fine-tuning. The authors evaluate OneVLA on VLN-CE, SimplerEnv, LIBERO, RoboCasa365, and real-world robots, reporting performance improvements over existing methods. Ablation studies further demonstrate the benefits of joint training and the progressive training strategy.

Critical strengths
  1. The paper addresses a practically relevant problem by unifying discrete navigation and continuous manipulation within one model. Its concrete implementation and evaluations across simulation benchmarks and real-world robotic platforms demonstrate value in system integration.
  2. Direct controlled comparisons support the benefits of joint training. Table 2b uses the same model framework and training steps, reporting improvements of 7.3, 9.3, and 5.5 percentage points on R2R, RxR, and SimplerEnv, respectively. These results provide evidence of mutual benefits from cross-task learning.
Critical weaknesses
  1. The unified action head’s independent contribution remains unclear. Table 2b supports joint training but does not isolate the benefit of head unification. A comparison against a shared backbone with two task-specific heads, using identical data and training settings, is missing. Table 8 cannot substitute for this control because it still involves differences in the backbone and training strategy.
  2. The independent contribution of CoT remains unclear. Table 9 attributes Stage 3 gains to CoT, but this stage also adds 10,000 training steps. An equal-step comparison with and without CoT, starting from the same Stage 2 checkpoint with otherwise identical settings, is needed to distinguish CoT benefits from additional training.
  3. Incomplete navigation metrics limit assessment of the claimed advantages. The paper reports only OSR, omitting NE, SR, and SPL on R2R, and these metrics plus nDTW on RxR. OSR measures whether a trajectory ever reaches the goal vicinity, but not final localization accuracy, correct stopping, path efficiency, or trajectory fidelity. OSR alone therefore cannot substantiate the claimed superiority in navigation performance.
Questions for authors

Please see Critical Weaknesses 1–3 for my main questions and concerns.

7 comments
AuthorsOpenReview ↗

Thanks and Response to Reviewer KPaZ (1/2)

We sincerely thank the reviewer for the thoughtful and constructive feedback. We greatly appreciate your recognition of our work’s “practical relevance” and “benefits of joint training”. Below, we address your main questions and concerns in detail.


Q1: Query about contribution of the unified action head.

Thanks for your question. We ran two comparisons to measure what a unified action head adds to joint training. In OneVLA, navigation and manipulation train the same DiT to turn visual-language features into actions.

(A) Head sharing under single-stage training. We used the single-stage setup from original Table 2b and trained two variants with separate navigation and manipulation heads. One used two full DiT-B heads. The other used two heads with the same combined parameter count as the unified head. All joint-training runs used the same setting.

Rebuttal Table 1. Head sharing under the original single-stage setting. Single-task and unified-head scores reproduce original Table 2b. The two separate-head rows are new controls.

Action headHead paramsR2R OSR ↑RxR OSR ↑SimplerEnv ↑
Single-task training (Navi. only / Mani. only, original Table 2b)1×48.833.640.7
Shared backbone + two task-specific heads2×53.439.643.5
Shared backbone + two task-specific heads (param-matched)1×52.738.542.9
Unified head (Ours, original Table 2b)1×56.142.946.2

On R2R, RxR, and SimplerEnv, joint training with a shared backbone and two full heads adds 4.6, 6.0, and 2.8 points over single-task training. Sharing the head adds another 2.7, 3.3, and 2.7 points, for total gains of 7.3, 9.3, and 5.5 points. The parameter-matched comparison checks whether the gain holds at the same total head size.

(B) Head sharing under the full training recipe. We then trained the unified-head and two-head models on the same 60K-step, three-stage schedule. Table 2 shows the results.

Rebuttal Table 2. Head sharing under the same three-stage schedule. OneVLA's scores are from the original paper. The separate-head row is a new control.

Action headR2R OSR ↑RxR OSR ↑SimplerEnv ↑
Shared backbone + two task-specific heads65.054.360.9
Unified head (OneVLA)68.658.264.5

The unified head improves R2R, RxR, and SimplerEnv by 3.6, 3.9, and 3.6 points, respectively. Original Table 8 also reports results from transferring the unified head to UniVLA. In original Table 7, removing inactive-dimension supervision reduced the mean from 63.8% to 42.8%. We will add both head comparisons to the revision and clarify the contribution of head sharing.


Q2: Query about the contribution of CoT.

Thanks for your question, At the same total training budget, our progressive training strategy improves R2R, RxR, and SimplerEnv by 12.5, 15.3, and 18.3 points in original Table 2a.

(A) Equal-step CoT comparison. To test whether CoT helps beyond extra training, we started from the same Stage 2 checkpoint and trained for another 10K steps in three settings: action-only, shuffled CoT, and aligned CoT as in the original Stage 3. We kept the trajectories, images, and training settings fixed across all three runs. Table 3 shows the results.

Rebuttal Table 3. CoT comparison from the same Stage 2 checkpoint. Stage 2 and aligned-CoT scores reproduce original Table 9. Action-only and shuffled-CoT scores are from the new controls.

SettingR2R OSR ↑RxR OSR ↑SimplerEnv ↑Avg. ↑
Stage 2 checkpoint (50K steps, original Table 9)63.753.459.158.7
+ 10K steps, w/o CoT (action-only)64.554.060.159.5
+ 10K steps, CoT with shuffled text63.953.259.458.8
+ 10K steps, CoT (Ours, Stage 3)68.658.264.563.8

An extra 10K steps of action-only training raised the average by 0.8 points. Aligned CoT added another 4.3 points, about 84% of the original 5.1-point Stage 3 gain. The shuffled run kept text supervision by assigning CoT to different observation–instruction pairs. It scored lower than the aligned run.

(B) Progress before Stage 3. We also evaluated the 40K and 45K checkpoints to see how much the model was still improving before Stage 3. Table 4 shows their scores alongside the 50K checkpoint.

Rebuttal Table 4. Late Stage 2 checkpoints. The 50K scores are from original Table 9. The 40K and 45K entries are new.

Cumulative steps40K45K50K (end of Stage 2)
R2R OSR ↑62.963.463.7
RxR OSR ↑52.653.153.4
SimplerEnv ↑58.358.859.1

Scores were leveling off by 50K steps, consistent with the small gain from continuing action-only training. We will add these checkpoint results and the equal-step CoT comparison to the revision.

AuthorsOpenReview ↗

Thanks and Response to Reviewer KPaZ (2/2)

Q3: Query about complete navigation metrics.

(A) Full navigation metrics. We measured NE, SR, and SPL on the R2R-CE and RxR-CE val-unseen splits and nDTW on RxR-CE. NE measures the final distance to the goal, SR measures successful stopping, SPL accounts for path efficiency, and nDTW measures route fidelity. Tables 5 and 6 report these metrics alongside the original OSR of 68.6% on R2R and 58.2% on RxR. The sources below each table identify the baseline results, including the StreamVLN version.

Rebuttal Table 5. Expanded R2R-CE evaluation. OneVLA's OSR is from original Table 1.

MethodParamsUnifiedNE ↓OSR ↑SR ↑SPL ↑
NaVILA7B✗5.2262.554.049.0
StreamVLN [ICRA’26]7B✗4.9864.256.951.9
UniVLA-Navi.* (OneVLA baseline)7B✗6.7147.138.234.5
OneVLA (Ours)3B✓4.7268.658.453.1

Sources · NaVILA Table I · StreamVLN v1 Table 1 · UniVLA Fig. 6 (R2R OSR only)

Rebuttal Table 6. Expanded RxR-CE evaluation. OSR scores are from original Table 1.

MethodParamsUnifiedNE ↓OSR ↑SR ↑SPL ↑nDTW ↑
NaVILA7B✗6.7753.649.344.058.8
StreamVLN [ICRA’26]7B✗6.2255.752.946.061.9
UniVLA-Navi.* (OneVLA baseline)7B✗8.3526.321.418.237.9
OneVLA (Ours)3B✓5.9858.253.746.862.6

Sources · NaVILA Table I · StreamVLN v1 Table 1

OneVLA leads StreamVLN v1 by 1.5 points in SR and 1.2 in SPL on R2R, and by 0.8 points on both metrics on RxR. OneVLA also has lower NE on both splits and higher nDTW on RxR.

(B) Effect of joint and progressive training. We also evaluated the navigation-only, single-stage joint-training, and full multi-stage models from original Tables 2a–2b. Table 7 reports R2R SR and SPL and RxR SR and nDTW alongside R2R OSR.

Rebuttal Table 7. Additional metrics for the original training ablations. The configurations and R2R OSR are from original Tables 2a–2b. SR, SPL, and nDTW are new.

SettingR2R OSR ↑R2R SR ↑R2R SPL ↑RxR SR ↑RxR nDTW ↑
Navi. only (original Table 2b)48.839.635.229.444.1
Joint training, single stage (original Table 2a/2b)56.146.341.838.151.7
Multi-stage training (OneVLA)68.658.453.153.762.6

SR, SPL, and nDTW improve with joint training and improve further with the multi-stage schedule.

We will add these ablations and the full navigation metrics to the revision.


We hope these responses address your concerns, and we would be glad to answer any further questions.

Sorry for getting u lately. According to the response, I still hold the concerns. The core reason is variances in SimplerEnv-Bridge is huge. 4 points did not mean the benefits. current submission lacks of comparisons on manipulation tasks.

For the real robot, I did not get the actual tasks for evaluation from this paper.

Comments on the paper

6 comments
AuthorsOpenReview ↗

To Reviewer knaj

Dear Reviewer knaj,

As the discussion period ends soon, we would greatly appreciate it if you could let us know whether our responses have addressed your concerns, or if there is anything else we can clarify.

Best regards, The Authors

Thanks for the response by the authors. The replies address some of my concerns. First, according to the results in Table 8, the advantage seems not very significant. Second, navigation and manipulation are evaluated separately and there is no coupled mobile manipulation evaluation. Therefore, I will keep my original score.

Thank you for the follow-up. We clarify the two remaining points below.


Q1: Query about significance of Rebuttal Table 8.

The small differences in Table 8 are the intended finding. Table 8 tests whether adding task-specific designs (longer navigation memory, separate action horizons, higher navigation resolution) improves over our shared configuration, and none of them does. For example, 31 history frames use 4× more visual tokens yet lower R2R SR from 58.4 to 57.4, and separate horizons lower it to 57.1 / 55.8. This shows that the shared configuration is already sufficient for both tasks and that the extra cost and complexity of task-specific modules bring no gain. The main improvements of OneVLA come from the unified action head and the training strategy (Rebuttal Tables 15–16 and original Tables 2a–2b), where the gains on R2R/RxR are 3.6–15.3 points.


Q2: Query about mobile manipulation evaluation.

We would like to clarify that our rebuttal already includes two coupled evaluations in which a single policy performs navigation and manipulation within the same episode:

EvaluationSettingUniVLAModular baselineOneVLA
RoboCasa365, standard start (original Table 6)50 tasks, base + arm in one episode9.9–28.4
RoboCasa365, long-range start (Rebuttal Table 10)robot starts farther from the target5.2–24.6
Real-world cross-room (Rebuttal Table 11)navigate between rooms, then manipulate with a fixed base30.040.062.5

In the real-world cross-room setting, the robot mobile manipulator: e.g., SLAMTEC base + arm receives one full instruction (e.g., "go to the kitchen and put the cup on the plate"), navigates between rooms, and then manipulates. OneVLA executes the whole task as a single policy without any task switching. In contrast, UniVLA needs an oracle switch between its task-specific heads, and the modular baseline (MapNav + Diffusion Policy) needs a hand-coded switch. OneVLA reaches 62.5% overall success, 22.5 points above the modular baseline and 32.5 points above UniVLA. We will move both coupled evaluations into the main paper, together with the platform details and per-stage results.


We hope these clarifications address your remaining concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.