- 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
-
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.
-
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
-
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.
-
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.
-
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
-
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.
-
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.
-
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
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.
| Architecture | Sampling | (cm) ↓ | (°) ↓ | SR (%) ↑ |
|---|---|---|---|---|
| (a) Unified tokens | Joint | 22.0 | 7.49 | 46.9 |
| (a) Unified tokens | Indep. + mask | 23.0 | 7.86 | 42.5 |
| (b) Separate experts | Joint | 19.7 | 6.06 | 54.4 |
| (c) Sequence concatenation | Joint | 22.1 | 6.78 | 60.0 |
| (d) Mixed-stream | Joint | 18.6 | 5.87 | 60.0 |
| (d) Mixed-stream, UniWAM-Base | Indep. | 17.1 | 5.30 | 63.1 |
(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.
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.
| Model | MAP-Data | Extra manip. data | (cm) ↓ | (°) ↓ | SR (%) ↑ |
|---|---|---|---|---|---|
| UniWAM-Base | ✗ | ✗ | 17.1 | 5.30 | 63.1 |
| UniWAM-Scaled w/o MAP-Data | ✗ | ✓ | 17.3 | 5.33 | 70.6 |
| UniWAM-Scaled w/o extra manip. data | ✓ | ✗ | 12.9 | 4.49 | 66.9 |
| UniWAM-Scaled | ✓ | ✓ | 13.2 | 4.56 | 80.0 |
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.