acceptodds
Desk rejected at ICLR 2027

UniWAM: Unified Mobile Manipulation via Mixed-Stream World-Action Modeling and Manipulation Anchor Pose Supervision

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

17 views

Abstract

Mobile manipulation requires precise navigation to a manipulation-ready pose followed by reliable object interaction. These two stages differ in action spaces and visual requirements, which complicates unified policy learning. In addition, collecting diverse real-world navigation data with explicit manipulation-ready pose supervision remains costly and difficult to scale. We introduce UniWAM, a unified mixed-stream world-action model with separate action encoders and output heads for navigation and manipulation, sharing a common backbone. This design supports joint representation learning on independently sampled navigation and manipulation data. UniWAM supports independent inference for either stream and batch-parallel inference for both. We further introduce Manipulation Anchor Pose (MAP) supervision for where to stop and how to orient for manipulation. An automated pipeline constructs MAP-Data from large-scale 3D scenes, yielding over 1.5 million episodes and 7,500 hours. MAP-Data provides per-frame target-object bounding boxes and image-plane MAP coordinates as auxiliary navigation supervision. Together with projected end-effector trajectories for manipulation, these prediction targets provide stream-specific image-plane supervision for action learning from egocentric observations. With large-scale MAP-Data, UniWAM outperforms the strongest external baselines on our MAP-Bench by 30.1% in position error and 44.0% in heading error. Across 24 real-robot tasks, UniWAM achieves leading results in MAP navigation and mobile manipulation, with competitive manipulation performance. We will release code, data, and benchmark.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Embodied AI, Mobile Manipulation, World-Action Models, Dataset and Benchmark

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_41ChOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

UniWAM addresses sequential mobile manipulation using independently sampled navigation and manipulation streams with separate visual inputs and action interfaces over shared video and action diffusion transformers. It combines flow-matching objectives with stream-specific image-plane supervision and introduces MAP-Data, containing over 1.5 million synthetic approach trajectories toward geometrically constructed manipulation anchor poses. Evaluation comprises MAP-Bench, zero-shot real-robot docking, and 24 real-robot tasks, with the scaled model reporting improved docking accuracy and 80.0% success on four mobile-manipulation tasks.

Critical strengths
  1. MAP-Data provides useful, transferable supervision for precise docking. For example, Tab. 7 evaluates models trained only on synthetic MAP-Data in unseen real rooms and with unseen target instances, and Tab. 8 further controls training budget and sampling mixture while increasing target diversity. Their results support the practical value of the dataset beyond matching synthetic benchmark labels.

  2. UniWAM-Scaled reaches 80.0% mobile-manipulation success versus 62.5% for π0.5, with reported 95% Wilson intervals of [73.1, 85.5] and [54.8, 69.6], respectively (Tab. 2 and 9). A single checkpoint serves the evaluated embodiments, and the authors explicitly acknowledge that fixed-base manipulation shows no clear success advantage over π0.5 and that scaling changes both data and compute, making the strengths and trade-offs reasonably transparent.

Critical weaknesses
  1. Equation (5) constructs each pose from a target-front normal and the boundary of navigable free space; Appendix B reports collision, visibility, and geometric filtering, but no quantitative validation of arm reachability, task-dependent feasibility, or manipulation success from those generated poses. MAP-Bench measures distance to this chosen reference, potentially penalizing other feasible poses, while the real-robot reference MAPs are manually recorded parking poses (Appendix E.3); therefore, even the valuable real docking transfer does not establish the manipulation feasibility of the synthetic labels, which is central to the dataset’s claimed purpose.

  2. Separate interfaces over a shared transformer with sampled dataset-specific updates are established in HPT (Wang et al., 2024, Secs. 3.1–3.2), as Appendix A acknowledges. UniWAM’s meaningful adaptation is applying this pattern within one mobile manipulator with separate views and world-action supervision. However, Table 3 restricts all alternative architectures to joint sampling, without a unified-token baseline using masked unavailable-action losses and the same independent source sampling and active/idle reweighting. The existing comparisons establish utility, but leave this simpler explanation open; the additional success gain from independent sampling is only 60.0% → 63.1%, and Appendix F.3 acknowledges overlapping success intervals without providing training-seed variation or uncertainty for docking errors.

  3. UniWAM-Scaled adds both MAP-Data and 402 manipulation tasks, while changing training from 20K steps on 32 GPUs to 30K steps on 64 GPUs (Table 6). UniWAM-Base reaches 63.1% mobile-manipulation success versus π0.5’s 62.5%, despite more accurate docking. The authors disclose these changes, and Tables 7–8 provide useful navigation-specific controls. But neither those controls nor the four mobile-manipulation tasks separates improved docking from a stronger manipulation policy as the source of the 80.0% result. This attribution gap materially limits the evidence for the proposed connection between MAP supervision and downstream manipulation reliability.

Questions for authors
  1. On a representative sample of automatically generated MAPs, what fraction permits successful manipulation for a specified task and embodiment? Please relate distance/heading error to manipulation success, including alternative feasible base poses.

  2. Does mixed-stream training outperform a unified-token baseline with the same backbone, auxiliary labels, independent source sampling, active/idle reweighting, and masked unavailable-action losses? Please include uncertainty for the key docking and success differences.

  3. Under a fixed training budget, how does mobile-manipulation success change when adding only MAP-Data, only the additional manipulation data, or both?

10 comments
AuthorsOpenReview ↗

Response to Reviewer 41Ch (1/3): Manipulation feasibility of MAP

We thank the reviewer for recognizing the transferable supervision of MAP-Data and the transparent reporting of results and trade-offs. Response 1 addresses the first weakness and question, with the fraction of feasible MAPs in (a), the agreement between synthetic MAPs and real working poses in (b), the relation between docking error and success in (c), and alternative base poses in (d). Response 2 addresses the second weakness and question, with the matched unified-token baseline in (a), the relation to HPT in (b), and the uncertainty of the key differences in (c). Response 3 addresses the third weakness and question, on the source of the 80.0% success, with a new fixed-budget data ablation (Table 3). Revisions to the manuscript are marked in blue.

Concern 1: On a representative sample of automatically generated MAPs, what fraction permits successful manipulation for a specified task and embodiment? Please relate distance/heading error to manipulation success, including alternative feasible base poses.

Response 1: We thank the reviewer for this question and answer it in four parts.

(a) Fraction of MAPs that permit manipulation. Our pipeline places every MAP under explicit embodiment constraints, which the original manuscript did not describe and which we have now added to Figure 2, Section 3.1, and Appendix B.2. A VLM selects the manipulation-facing side of the target from a top-down view of the scene. The MAP is then placed so that the front of the chassis keeps a 10 cm safety clearance from the target, and a target is kept only if the 62.6 cm reach of the PiPER arm covers the distance from the arm base to the target's semantic center. Every approach trajectory is also checked against the 70.2 × 61.0 cm base footprint. By construction, all MAPs in MAP-Data therefore place the target within the nominal reach of the arm. This is a geometric reachability guarantee rather than a guarantee of task success. A per-MAP manipulation success rate would require physically interactive object assets and a manipulation policy for every target, which scanned and reconstructed 3D scenes do not provide. We therefore measure manipulation readiness on real robots.

(b) Agreement between synthetic MAPs and real working poses. The real-robot reference MAPs are the working poses at which the manipulation demonstrations were collected, chosen by operators to face the target at a manipulable distance (Appendix E.3). A UniWAM trained only on synthetic MAP-Data, without any real-robot data, stops on average 18.7 cm from these real working poses in unseen rooms (Table 8), close to the 15.1 cm of UniWAM-Base trained on real demonstrations. The synthetic MAP definition therefore agrees with the poses that human operators choose for manipulation.

(c) Relation between docking error and manipulation success. In the controlled ablation of Table 3, the architecture, manipulation data, and training budget are held fixed. Adding MAP-Data reduces the mean docking error from 17.3 cm to 13.2 cm and 5.33° to 4.56°, while increasing end-to-end success from 70.6% to 80.0% over 160 real-robot trials. The per-task results of UniWAM-Scaled in Table 11 show position errors of 12.8–13.8 cm, heading errors of 4.37–4.83°, and success rates of 75.0–85.0% across the four tasks. Task-level success is not strictly monotonic in docking error, because manipulation difficulty also differs across tasks. With the manipulation capability held fixed, however, more accurate docking is accompanied by higher end-to-end success.

(d) Alternative feasible base poses. We agree that a target can often be manipulated from several base poses. Among them, MAP selects a geometrically favorable one. The base faces the manipulation-facing edge along its normal at the safety clearance, so the target is centered in front of the robot, within the reach of the arm and the view of the head camera. Real-robot results support this choice. Besides the controlled result in (c), removing navigation auxiliary supervision (Table 2) leaves fixed-base Manipulation success nearly unchanged, 43.6% versus 44.3%, but increases the Mobile Manipulation docking error from 17.1 cm to 19.9 cm and lowers success from 63.1% to 56.9%. In both controlled comparisons, stopping closer to the MAP leads to more reliable manipulation. MAP also serves as a consistent reference, since the real demonstrations were collected at the reference MAP and a vision-based policy performs best from a similar camera view and arm–target configuration. The MAP-Bench error therefore measures deviation from this reference, not infeasibility of an alternative pose.

AuthorsOpenReview ↗

Response to Reviewer 41Ch (2/3): Matched unified-token baseline and uncertainty

Concern 2: Does mixed-stream training outperform a unified-token baseline with the same backbone, auxiliary labels, independent source sampling, active/idle reweighting, and masked unavailable-action losses? Please include uncertainty for the key docking and success differences.

Response 2: We thank the reviewer for this suggestion. We added this baseline and evaluated the statistical reliability of the comparison.

Table 4 (revised). Comparison of unified architectures and sampling strategies on the real-robot Mobile Manipulation tasks.

ArchitectureSamplingEposE_{\mathrm{pos}} (cm) ↓EheadE_{\mathrm{head}} (°) ↓SR (%) ↑
(a) Unified tokensJoint22.07.4946.9
(a) Unified tokensIndep. + mask23.07.8642.5
(b) Separate expertsJoint19.76.0654.4
(c) Sequence concatenationJoint22.16.7860.0
(d) Mixed-streamJoint18.65.8760.0
(d) Mixed-stream, UniWAM-BaseIndep.17.15.3063.1

(a) Comparison under matched training conditions. The new Unified Tokens baseline uses the same backbone, auxiliary labels, independent source sampling, active/idle reweighting, and training budget as Mixed-stream, with unavailable action dimensions masked in the loss, denoted Indep. + mask in Table 4. With unified tokens, independent sampling amounts to masking the navigation dimensions in most manipulation samples and the manipulation dimensions in most navigation samples. As shown in the revised Table 4, Mixed-stream reduces docking error from 23.0 cm to 17.1 cm and 7.86° to 5.30°, and improves end-to-end success from 42.5% to 63.1% (+20.6 percentage points). Under joint sampling, the advantage is similar, 60.0% versus 46.9%. Independent sampling does not help Unified Tokens, likely because each shared token then receives supervision on only part of its dimensions in most samples. Since a unified token couples two physically different actions, masking part of it has no clear physical meaning, and we conjecture that this hinders learning its more complex representation. The gain of Mixed-stream therefore comes from the stream-specific action interfaces rather than from independent sampling or loss masking alone.

(b) Relation to HPT. HPT (Wang et al., 2024) uses a shared trunk with embodiment-specific interfaces across different robots. UniWAM applies stream-specific interfaces to the navigation and manipulation phases of one mobile manipulator, combined with independent stream sampling, world-action modeling, and image-plane supervision, as discussed in Appendix A.

(c) Statistical analysis. Each model is evaluated on 160 real-robot trials (40 per task). Compared with Unified Tokens, Mixed-stream reduces position error by 5.90 cm (95% CI: [4.01, 7.77] cm) and heading error by 2.56° (95% CI: [1.81, 3.31]°), and improves success by 20.6 percentage points (95% CI: [9.7, 30.9]). Training both models with three independent random seeds, Mixed-stream outperforms Unified Tokens in every seed. It reduces docking position error by 5.93 ± 0.65 cm and heading error by 2.57 ± 0.45°, while improving end-to-end success by 20.21 ± 5.01 percentage points (mean ± standard deviation across seeds). Details are given in Appendix F.3 and F.5.

AuthorsOpenReview ↗

Response to Reviewer 41Ch (3/3): Fixed-budget data ablation

Concern 3: Under a fixed training budget, how does mobile-manipulation success change when adding only MAP-Data, only the additional manipulation data, or both?

Response 3: We thank the reviewer for this suggestion and added the requested ablation as the revised Table 3. UniWAM-Scaled and two variants that each remove one data source share the architecture, initialization, budget of 30K steps on 64 GPUs, and navigation and manipulation sampling ratio, so they differ only in whether MAP-Data and the 402 additional manipulation tasks are used. UniWAM-Base, trained with a smaller budget, is listed for reference.

With the additional manipulation data fixed, adding MAP-Data reduces the docking error from 17.3 cm / 5.33° to 13.2 cm / 4.56° and raises success from 70.6% to 80.0%. With MAP-Data fixed, adding the manipulation data raises success from 66.9% to 80.0%, while docking changes little, from 12.9 cm / 4.49° to 13.2 cm / 4.56°. MAP-Data therefore mainly improves where the robot stops, and the manipulation data mainly improves what it does from there. Both contribute to the 80.0% of UniWAM-Scaled, and MAP-Data improves end-to-end success independently of the additional manipulation data and compute.

Table 3 (new). Ablation of MAP-Data and the additional manipulation data on the real-robot Mobile Manipulation tasks.

ModelMAP-DataExtra manip. dataEposE_{\mathrm{pos}} (cm) ↓EheadE_{\mathrm{head}} (°) ↓SR (%) ↑
UniWAM-Base✗✗17.15.3063.1
UniWAM-Scaled w/o MAP-Data✗✓17.35.3370.6
UniWAM-Scaled w/o extra manip. data✓✗12.94.4966.9
UniWAM-Scaled✓✓13.24.5680.0
Reviewer_zDUhOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes UniWAM, a unified mobile manipulation framework that models navigation and manipulation as two independently sampled streams with separate action interfaces and a shared world-action backbone. The authors also introduce Manipulation Anchor Pose (MAP) supervision and a large-scale synthetic dataset, MAP-Data, to improve manipulation-oriented navigation. Experiments on MAP-Bench and 24 real-robot tasks demonstrate improved navigation accuracy and mobile manipulation success rates.

Critical strengths
  1. Practical problem and scalable data generation. The paper addresses the important challenge of connecting navigation and manipulation. The large-scale MAP-Data provides explicit supervision for navigation terminal poses and offers a potentially useful resource for mobile manipulation research.

  2. Extensive experimental evaluation. The authors evaluate the method on multiple robot platforms and tasks, with comparisons against strong VLA baselines and several architecture and supervision ablations. The improvements in navigation accuracy and mobile manipulation success are encouraging.

Critical weaknesses
  1. The manipulation awareness of MAP is not sufficiently justified. MAP is generated mainly through geometric heuristics, including object-facing edges and navigable-space boundaries, without explicitly considering robot arm reachability, manipulation feasibility, or downstream policy success. Therefore, it is unclear whether the generated poses are genuinely manipulation-ready or simply geometrically reasonable navigation endpoints. The current evaluation mainly measures agreement with these generated MAP labels, which does not independently establish their suitability for manipulation.

  2. The necessity of the shared backbone is insufficiently demonstrated. Although UniWAM is presented as a unified model, navigation and manipulation use separate observations, action interfaces, and independently sampled training examples, with no direct cross-stream attention. The paper does not compare this design against two independently trained navigation and manipulation policies under matched data and compute budgets. Consequently, it remains unclear whether sharing the backbone enables meaningful cross-task knowledge transfer or merely provides a convenient parameter-sharing implementation.

Questions for authors

Q1. How do the authors validate that the geometrically generated MAP poses are actually suitable for downstream manipulation? Could MAP be evaluated against manipulation success, arm reachability, or policy-conditioned optimal base poses rather than primarily against the generated geometric labels?

Q2. What are the benefits of sharing the backbone compared with training two separate navigation and manipulation models under matched data and compute budgets? Is there evidence of positive transfer between the two streams?

4 comments
AuthorsOpenReview ↗

Response to Reviewer zDUh (1/2): Manipulation readiness of MAP

We thank the reviewer for recognizing the scalable data generation and the extensive real-robot evaluation. Response 1 addresses the first weakness and Q1, with arm reachability in (a), manipulation success in (b), and policy-conditioned base poses in (c). Response 2 addresses the second weakness and Q2, with the comparison against two separately trained models in (a), evidence of joint learning in (b) and (c), and the training and deployment benefits in (d). Revisions to the manuscript are marked in blue.

Concern 1: How do the authors validate that geometrically generated MAP poses are suitable for downstream manipulation, beyond agreement with the generated labels? Could their suitability be assessed through arm reachability, manipulation success, or policy-conditioned optimal base poses?

Response 1: We thank the reviewer for this important question.

(a) MAP construction includes embodiment constraints. MAP is not a purely geometric navigation endpoint. Our pipeline places each MAP so that the front of the chassis keeps a 10 cm safety clearance from the target and the target remains within the 62.6 cm reach of the PiPER arm, and it checks every approach trajectory against the 70.2 × 61.0 cm footprint of the TRACER 2.0 base. Every MAP therefore places the target within the nominal reach of the arm. The original manuscript focused on the geometric construction and omitted these constraints, and we have now made them explicit in Figure 2, Section 3.1, and Appendix B.2.

(b) Real-robot experiments validate the downstream benefit of MAP-Data. In the new controlled ablation (revised Table 3), the manipulation data, architecture, and training budget are fixed. Adding MAP-Data reduces docking error from 17.3 cm to 13.2 cm and 5.33° to 4.56°, and raises end-to-end success from 70.6% to 80.0%. MAP supervision therefore improves the downstream manipulation on real robots, not only the agreement with the generated labels.

(c) MAP targets scalable and standardized supervision. N2M (Chai et al., 2025), Mobi-π (Yang et al., 2025), and GBPP (Chen et al., 2026b) optimize base poses for a specific robot or policy through policy rollouts or simulation feedback, which is costly to scale across diverse environments. MAP instead uses geometric construction with configurable embodiment constraints on large-scale scanned 3D scenes, which enables over 1.5M approach trajectories (Table 6) with zero-shot real-robot transfer (Table 8). Vision-based VLA and WAM policies need approach trajectories that start from diverse poses but end in a consistent pose relative to the target, so that manipulation always begins from a similar view. MAP-Data provides such trajectories at scale, since every approach starts from a different observation pose and ends at the same target-facing heading and working distance.

AuthorsOpenReview ↗

Response to Reviewer zDUh (2/2): Benefit of the shared backbone

Concern 2: What are the benefits of sharing the backbone compared with training two separate navigation and manipulation models under matched data and compute budgets? Is there evidence of positive transfer between the two streams?

Response 2: We thank the reviewer for this suggestion. We added a direct comparison with two independently trained WAMs in the revised Table 4.

(a) Comparison with two independent WAMs. We train a navigation WAM and a manipulation WAM separately on the same UniWAM-Base data and execute them sequentially. Each receives the full UniWAM-Base budget so that neither is undertrained, and the pair therefore uses twice the compute of UniWAM-Base. The independent WAMs achieve similar docking errors (16.8 cm / 5.20°) to UniWAM-Base (17.1 cm / 5.30°), but a lower end-to-end success rate, 58.8% compared with 63.1%. UniWAM thus completes the task more often with half the training compute, which indicates that the shared backbone benefits the complete task.

(b) Joint learning through the shared backbone. As shown in Figure 4(d), UniWAM uses separate visual inputs, action encoders, output heads, and learned tokens for the two streams, while sharing the Video DiT and Action DiT. Although there is no cross-row attention, both streams update the shared parameters, which allows joint representation learning from navigation and manipulation data (Section 3.2).

(c) Further evidence from architecture ablations. In Table 4, Joint denotes joint whole-body sampling, where each sampled window supervises both streams, and Indep. denotes independent stream sampling. Under the same data, budget, and joint sampling, Mixed-stream outperforms separate Action DiTs (Figure 4(b)), reducing docking errors from 19.7 cm / 6.06° to 18.6 cm / 5.87° and raising success from 54.4% to 60.0%.

(d) Lower training and deployment cost. One training run covers both streams, whereas two separate models need two runs and, in our comparison, twice the compute. Independent stream sampling also lets navigation-only and manipulation-only data be used directly. At deployment, a single checkpoint serves both streams without duplicating parameters. Batch packing runs the two streams in one forward pass, and together with our other inference optimizations it reduces dual-stream inference latency by more than half, from about 220 ms to below 60 ms (Section 3.2 and Appendix D.1).

In short, the unified model learns a joint representation that improves task success, and it is also cheaper to train and deploy, with fast batch-parallel inference.

Table 4 (revised). Comparison of unified architectures and sampling strategies on the real-robot Mobile Manipulation tasks.

ArchitectureSamplingEposE_{\mathrm{pos}} (cm) ↓EheadE_{\mathrm{head}} (°) ↓SR (%) ↑
(a) Unified tokensJoint22.07.4946.9
(a) Unified tokensIndep. + mask23.07.8642.5
(b) Separate expertsJoint19.76.0654.4
(c) Sequence concatenationJoint22.16.7860.0
(d) Mixed-streamJoint18.65.8760.0
(d) Mixed-stream, UniWAM-BaseIndep.17.15.3063.1
Independent WAMsNav. + Manip.16.85.2058.8
AuthorsOpenReview ↗

Follow-up on MAP Manipulation Readiness and Shared Backbone

Dear Reviewer zDUh,

Thank you again for your thoughtful comments. We would like to kindly follow up on our detailed responses.

Regarding MAP's manipulation readiness, we clarified the embodiment constraints and provided a controlled real-robot ablation showing that MAP-Data improves end-to-end success from 70.6% to 80.0%. Regarding the shared backbone, we added a comparison with two independently trained WAMs (Table 4), showing higher success (63.1% vs. 58.8%) with half the training compute.

We would be grateful to know whether these new results and clarifications have addressed your main concerns, or whether further clarification would be helpful.

As the discussion period is nearing its end, we sincerely hope to hear your feedback. Thank you very much for your time and consideration.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Rating Reconsideration

Dear Reviewer zDUh,

Thank you again for your thoughtful review and constructive feedback. We have addressed your concerns with additional evidence of MAP's downstream manipulation benefits, a direct comparison against independently trained WAMs, and further clarifications in the revised manuscript.

We would greatly appreciate your feedback on whether these additions have resolved your main concerns. If so, would you consider raising your rating in light of the revised manuscript?

Thank you very much for your time and consideration.

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

This paper addresses the different visual and action requirements of navigation and manipulation. UniWAM uses separate control streams with shared backbones and independent sampling. It introduces MAP supervision and over 1.5 million synthetic navigation episodes for precise docking. Results on MAP-Bench and 24 real-robot tasks show improved docking and mobile manipulation performance, with competitive manipulation results.

Critical strengths
  1. Navigation and manipulation need different visual inputs and actions. Using separate streams with a shared backbone is a reasonable way to address these differences.
  2. MAP-Data provides a large set of navigation trajectories with labels for where to stop and how to face the target. Experiments with different dataset sizes and zero-shot tests on real robots provide evidence for its value.
Critical weaknesses

The scaled model uses both MAP-Data and extra manipulation data. This makes it hard to tell whether better navigation from MAP-Data alone helps the robot complete the full mobile manipulation task.

Questions for authors
  1. A comparison with and without MAP-Data would help show its effect on mobile manipulation success. Both settings should use the same manipulation data and training budget.
  2. More detail on robot size and safety margins would help explain how MAP positions are chosen.
3 comments
AuthorsOpenReview ↗

Response to Reviewer eNmi: MAP-Data ablation and embodiment constraints

We thank the reviewer for recognizing the stream design of UniWAM and the transferability of MAP-Data. Response 1 addresses the weakness and the first question, on whether MAP-Data alone improves mobile manipulation success, with a new controlled ablation (Table 3). Response 2 addresses the second question by describing how arm reach, safety clearance, and base footprint constrain MAP construction. Revisions to the manuscript are marked in blue.

Concern 1: The scaled model uses both MAP-Data and additional manipulation data, making it difficult to isolate the contribution of MAP-Data to end-to-end mobile manipulation performance. How does MAP-Data independently affect mobile manipulation success when the manipulation data and training budget are kept identical?

Response 1: We thank the reviewer for this suggestion. We added a controlled ablation in the revised Table 3 (Section 4.3). We compare models trained with and without MAP-Data while keeping the architecture, initialization, manipulation data, training budget, and navigation/manipulation sampling allocation fixed. The evaluation follows Table 2, with four mobile manipulation tasks and 40 trials per task.

With identical manipulation data and training budget, adding MAP-Data reduces the docking position error from 17.3 cm to 13.2 cm (23.7%) and the heading error from 5.33° to 4.56° (14.4%), and raises end-to-end success from 70.6% to 80.0% (+9.4 percentage points).

This gain comes from the hand-over between the two stages. A vision-based VLA or WAM policy starts manipulating from wherever the base stops, so the stopping pose sets both its camera view and the arm–target geometry. MAP supervision makes the base stop more accurately and more consistently, so manipulation starts from conditions closer to those in the demonstrations. The 9.4-point gain with the manipulation data fixed shows this effect on real robots.

Table 3 (new). Ablation of MAP-Data and the additional manipulation data on the real-robot Mobile Manipulation tasks.

ModelMAP-DataExtra manip. dataEposE_{\mathrm{pos}} (cm) ↓EheadE_{\mathrm{head}} (°) ↓SR (%) ↑
UniWAM-Base✗✗17.15.3063.1
UniWAM-Scaled w/o MAP-Data✗✓17.35.3370.6
UniWAM-Scaled w/o extra manip. data✓✗12.94.4966.9
UniWAM-Scaled✓✓13.24.5680.0

 

Concern 2: How are the robot's physical dimensions, safety clearances, and manipulator workspace incorporated into the selection of MAP positions?

Response 2: We thank the reviewer for pointing out this important detail. Our real data pipeline already applies embodiment constraints, but the original manuscript focused on the geometric construction of MAP and did not describe them. We agree that these constraints are essential for manipulation-ready navigation, and we have added them to Figure 2, Section 3.1, and Appendix B.2.

The pipeline uses three embodiment parameters.

  • Arm reach rr, the nominal maximum reach of the manipulator.
  • Safety clearance cc, the distance kept between the front of the chassis and the target or any obstacle.
  • Base footprint L×WL \times W, the length and width of the mobile base.

The first two parameters constrain MAP generation (stage S2 in Figure 2). The MAP is placed so that the front of the chassis stops at distance cc from the target's manipulation-facing edge, and a target is kept only if rr exceeds the sum of the distance from the arm base to the front of the chassis, the clearance cc, and the distance from the manipulation-facing edge to the target's semantic center. The footprint constrains trajectory synthesis (stage S3), where every trajectory is checked so that the whole base, not only its center, stays collision free in narrow passages. For the reference Cobot Magic, rr = 62.6 cm (PiPER), cc = 10 cm, and L×WL \times W = 70.2 × 61.0 cm (TRACER 2.0).

These parameters are configurable for other embodiments, while MAP-Bench fixes the reference values for consistent evaluation.

AuthorsOpenReview ↗

Follow-up on MAP-Data Ablation and Embodiment Constraints

Dear Reviewer eNmi,

Thank you again for your constructive feedback. We would like to kindly follow up on our rebuttal.

To address your main concerns, we added a controlled ablation (Table 3) showing that MAP-Data independently improves end-to-end mobile manipulation success from 70.6% to 80.0% under the same manipulation data and training budget. We also clarified the embodiment constraints, including arm reach, safety clearance, and base footprint, in the revised manuscript.

We would greatly appreciate hearing whether these additions have adequately addressed your concerns, or if any questions remain.

As the discussion period is approaching its end, we sincerely look forward to your feedback. Thank you again for your time and consideration.

AuthorsOpenReview ↗

Follow-up on Rebuttal and Rating Reconsideration

Dear Reviewer eNmi,

Thank you again for your thoughtful review and constructive feedback. We have addressed your concerns through a controlled MAP-Data ablation, additional experimental evidence, and clarifications in the revised manuscript.

We would greatly appreciate your feedback on whether these additions have resolved your main concerns. If so, would you consider raising your rating in light of the revised manuscript?

Thank you very much for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions and Additional Experiments

Dear Area Chair and Reviewers,

Thank you for your thoughtful feedback. We have posted our point-by-point responses and revised the manuscript, with changes highlighted in blue. To facilitate the discussion, we summarize the main additions below.

  1. Isolating the contribution of MAP-Data (eNmi, zDUh, 41Ch; Table 3). With the architecture, initialization, manipulation data, and training budget held fixed, adding MAP-Data reduces docking position error from 17.3 cm to 13.2 cm and heading error from 5.33° to 4.56°, while increasing end-to-end mobile manipulation success from 70.6% to 80.0%.

  2. Strengthening the architecture comparisons (41Ch, zDUh; Table 4 and Appendices F.3/F.5). We added a unified-token baseline with matched independent sampling, loss masking, auxiliary supervision, and training budget. Mixed-stream achieves 63.1% success versus 42.5%, with consistent improvements across three training seeds. We also added two independently trained WAMs: they achieve 58.8% success using twice the total training compute of UniWAM-Base, compared with 63.1% for the shared model.

  3. Clarifying MAP construction and its scope (eNmi, zDUh, 41Ch; Figure 2, Section 3.1, and Appendix B.2). We now explicitly describe the arm-reach, safety-clearance, and base-footprint constraints used in data generation. We distinguish nominal geometric reachability from guaranteed manipulation success, and clarify that MAP-Bench measures deviation from a reference pose rather than the infeasibility of alternative poses. The controlled real-robot ablation above provides evidence of the downstream benefit of MAP-Data.

We would appreciate your feedback on whether these additions address your concerns, particularly regarding MAP’s downstream utility and the mixed-stream design. If any questions remain, we would be happy to provide further clarification during the discussion period.

Thank you for your time and consideration.

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

Related papers

Discussion (0)

Sign in to comment.