{"id":"da68352a-7cef-404f-b608-d5e9586b7694","arxiv_id":"2509.01658","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"MoTo turns existing fixed-base manipulation models into mobile manipulators by using VLM-picked contact keypoints and trajectory optimization to find docking points, with no training of MoTo itself.","lead":"A new plug-in module, MoTo, lets off-the-shelf fixed-base manipulation models operate on mobile robots by computing where the base should stop before grasping. It uses a vision-language model to select contact points on the target object and the robot arm, then optimizes base and arm trajectories to bring those points together.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"MoTo's zero-shot claim is contradicted by its own fine-tuning of backbone models on mobile manipulation expert data (App. A.1/A.3), so the reported gains do not isolate MoTo.","rationale":"The reader's formal weakest_assumption concerns the scope of tasks that decompose into 'navigate then fixed-base manipulate' and the full-observability assumption, which the paper itself acknowledges in Section 7. That is a genuine limitation but not an internal inconsistency. The more load-bearing concern is the gap between the zero-shot/no-expert-data claim and the fine-tuning described in Appendices A.1 and A.3. This is not an attack on the authors: the MoTo module itself may still be zero-shot and the cost terms in Section 4.3 are internally coherent. But the experimental evidence as presented cannot separate MoTo's contribution from the effect of fine-tuning the manipulation backbones on mobile manipulation demonstrations. A frozen-backbone evaluation would settle this cleanly. The reader's conditional verdict remains appropriate; I would keep it unchanged, with the explicit condition that the zero-shot claim be tested without fine-tuning the off-the-shelf manipulation models.","tokens_in":715,"tokens_out":740,"duration_ms":55578,"concrete_test":"Evaluate MoTo on OVMM and the three real-world tasks twice per backbone: (i) with off-the-shelf released checkpoints of OpenVLA, RDT-1B, and iDP3 used exactly as-is (no fine-tuning, no 20k/50 demonstrations from App. A.3), and (ii) with the fine-tuned checkpoints. Report Overall SR for both. If condition (i) is not close to condition (ii) for MoTo and still exceeds the L3MVN/fixed-base baseline, the zero-shot claim fails and the Table 1/Fig. 3 gains should be attributed to the fine-tuning rather than to MoTo.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that MoTo is a zero-shot plug-in: any off-the-shelf fixed-base manipulation model becomes mobile without any requirement on mobile manipulation expert data. The paper's own experiments do not test this. Appendix A.1 says the authors used an OVMM heuristic baseline to collect manipulation expert trajectories and fine-tune off-the-shelf manipulation foundation models. Appendix A.3 reports collecting 20k expert demonstrations during mobile manipulation to fine-tune OpenVLA with LoRA, and 50 expert demonstrations per task in a real kitchen to fine-tune RDT-1B and iDP3. These are exactly the mobile manipulation expert data the abstract says are not required. Consequently, the OVMM gains in Table 1 (e.g., OpenVLA w/ MoTo Overall SR 20.64%) and real-world gains in Figure 3 are confounded: they may reflect the fine-tuning of the manipulation backbone rather than MoTo's docking-point search. Without a frozen-backbone condition, the central zero-shot plug-in claim is not empirically established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MoTo, a plug-in module that converts fixed-base manipulation policies into mobile manipulation policies by (1) grounding an instruction in a 3D scene graph, (2) using a VLM to propose and select interaction keypoints on the target object and the robot arm under multi-view voting, and (3) solving a trajectory optimization that moves the base and arm so that the two keypoints align while respecting collision, smoothness, and arm-margin costs. The authors claim that MoTo provides zero-shot mobile manipulation for any off-the-shelf fixed-base manipulation model without requiring mobile manipulation expert data, and they report improved success rates on the OVMM benchmark and on a real dual-arm platform with several manipulation backbones.","tokens_in":15644,"tokens_out":5379,"duration_ms":58579,"significance":"If the zero-shot plug-in claim held, this would be a useful contribution: it would allow the rapidly improving class of fixed-base VLA and diffusion policies to be transferred to mobile platforms without training new mobile policies. The paper has clear strengths: the system is modular, the multi-view keypoint voting is a sensible way to reduce VLM grounding noise, the pipeline is described concretely, and the authors evaluate with several manipulation backbones in both simulation and the real world. They also state limitations honestly in Section 7. However, as detailed below, the experimental setup does not actually test the headline zero-shot claim, and the quantitative evidence is weakened by absent variance reporting and by the omission of the closest prior baseline. The core idea is defensible, but the claims need to be re-scoped or the experiments need to be extended.","major_comments":[{"comment":"The central claim that MoTo works zero-shot 'without any requirement on mobile manipulation expert data' is not tested by the paper's experiments. Appendix A.1 states that the authors collected manipulation expert trajectories using an OVMM-heuristic baseline and fine-tuned off-the-shelf manipulation foundation models. Appendix A.3 reports collecting 20k demonstrations to fine-tune OpenVLA with LoRA, and 50 expert demonstrations per task to fine-tune RDT-1B and iDP3 in a real kitchen. These are exactly the kind of mobile-manipulation expert data the abstract says are unnecessary. Consequently, the OpenVLA-with-MoTo gains in Table 1 and the real-world gains in Figure 3 are confounded: they may be due to fine-tuning the manipulation backbone rather than to MoTo's docking-point search. Please either evaluate frozen manipulation backbones without any fine-tuning, or revise the zero-shot/no-e","section":"Abstract, §5, App. A.1/A.3"},{"comment":"The main quantitative claims lack uncertainty estimates. Table 1 reports single point estimates for each success rate, and Figure 3 reports only 10 real-world trials per task with no standard deviation, confidence interval, or per-trial breakdown. An improvement of 2.68% on OVMM (OpenVLA w/ MoTo vs. UniTeam in Table 1) and 16.67% in the real world can easily lie within sampling noise at this sample size. Please report the number of OVMM validation episodes, multiple evaluation seeds, per-task success counts, error bars for the real-robot results, and, where possible, a significance test before claiming state-of-the-art performance.","section":"§5.1–5.3, Table 1, Fig. 3"},{"comment":"The comparison omits the closest published baseline. MoTo is closely related to MoManipVLA [3], a base-waypoint-optimization method for transferring VLA policies to mobile manipulation, and to IALP [27], but neither appears in Table 1. Since the manuscript claims state-of-the-art performance and MoTo's contribution is precisely a different way of choosing docking points, the evaluation must include those methods. Without them, the reported margins do not support the SOTA claim.","section":"§5.1, Table 1"}],"minor_comments":[{"comment":"The column header appears garbled: 'A verage SR Step' mixes two columns, and numbers such as 1132.5 cannot be success rates. Separate the 'Average SR' column from the 'Step' column and label the latter with its units (presumably episode length).","section":"Table 1"},{"comment":"The real-world results are shown only as a dot plot. Please add a companion table with mean success rate, task completion rate, number of trials, and per-task counts, since the dots are hard to read and the claimed 16.67% improvement is not accompanied by numerical values in the text.","section":"Figure 3"},{"comment":"Several hyperparameters—τ in Eq. (5), ε0 in Eq. (8), topV, rmin/rmax in Eq. (10), Nq, and the dual-annealing stopping threshold—are never given concrete values. A reproducibility table with these settings would be valuable, as would a brief sensitivity study for τ and ε0.","section":"§4.2–4.3"},{"comment":"The equation numbering is inconsistent: the paper refers to Eq. (1) as the overall objective, then to Eq. (3) and Eq. (4) and Eq. (7) for components. Algorithm 1 says 'Use Eq. (4)' but also calls the cost function O. Please align the numbering and state explicitly what O is in Algorithm 1.","section":"§4.1–4.3, Appendix C"},{"comment":"Minor typos and formatting issues: 'Inter view' in Figure 1 should be 'multi-view'; 'Double Annealing' in Appendix C should be 'Dual Annealing'; reference [37] contains 'V oxPoser' with a stray space; and the abstract's 'interaction keypoints framework' should be 'interaction-keypoint framework'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper is from the same research group as MoManipVLA [3] and does not compare against it, which is a particular concern given that MoManipVLA is arguably the closest prior work. The appendix descriptions of fine-tuning on mobile-manipulation demonstrations directly undercut the abstract's no-expert-data claim; this is a claims-consistency issue that the editor should weight heavily. The paper is not beyond repair: adding frozen-backbone evaluations, variance reporting, and the missing baselines would substantially strengthen it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is genuinely useful: instead of retraining a whole mobile manipulation policy, generate task-relevant interaction keypoints on the object and the arm via VLM multi-view voting, then optimize the base and arm trajectories to bring those keypoints together. That is a clean, modular way to extend fixed-base manipulation models to mobile settings, and it is new relative to ReKep (static, single-view) and MoManipVLA (which lacks part-level keypoints). The writing is clear, the optimization formulation is coherent, and the limitation section is refreshingly direct. The paper deserves credit for that.\n\nThe soft spots are real, though. The central claim — zero-shot plug-in, no mobile manipulation expert data — is contradicted by the paper's own appendix. A.1 says they used a heuristic baseline to collect manipulation expert trajectories to fine-tune off-the-shelf models. A.3 reports 20k demonstrations to fine-tune OpenVLA, and 50 per task for RDT-1B and iDP3. Those are exactly the mobile manipulation expert data the abstract says are not required. So the OVMM gain of 2.68% is confounded: it may come from the fine-tuned backbone rather than from MoTo's docking search. I would not call this fatal — the modular pipeline is still a sensible contribution — but the paper must include a frozen-backbone condition or explicitly separate the gains. Without that, the headline claim is not established.\n\nOther problems are more conventional. The main table has no error bars. The real-world experiments are 10 trials per task, which is thin for a 16.67% claim. The most directly related baseline, MoManipVLA (from the same group), is omitted from the comparison. The task scope is also limited to navigation-then-fixed-manipulation; the authors admit they cannot handle whole-body coordination like door opening, so the 'general mobile manipulation' framing oversells the actual coverage.\n\nOn balance, the method is worth taking seriously. The keypoint voting is a real idea, the optimization is well-specified, and the limitations are stated in-scope. But the paper needs a revision: frozen-backbone experiments, MoManipVLA baseline, variance reporting, and a more careful zero-shot claim. I would send it to peer review — the community can benefit from the idea even if the current evidence is modest.","headline":"A plausible modular docking-point method that is undermined by its own fine-tuned backbones: the zero-shot claim is asserted, not tested, and the missing MoManipVLA baseline plus no variance reporting make the empirical case weaker than the writing suggests.","tokens_in":803,"tokens_out":2737,"would_cite":false,"duration_ms":35385,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A plug-in that turns any fixed-base manipulation model into a mobile one, zero-shot, without mobile-manipulation training data.","keywords":["mobile manipulation","zero-shot learning","vision-language models","interaction keypoints","docking-point navigation","trajectory optimization","manipulation foundation models","scene graphs"],"falsifier":"Place the target object behind a partial occlusion so it is visible only at low resolution in each camera view; MoTo's keypoint proposal and voting should degrade sharply. If a single low-resolution view still yields successful docking and manipulation, the full-visibility assumption is not load-bearing.","tokens_in":15294,"feed_emoji":"🤖","tokens_out":6982,"duration_ms":63765,"temperature":0.7,"pith_summary":"Mobile manipulation usually needs large amounts of demonstration data, but fixed-base manipulation models already generalize well from large training sets. MoTo is a plug-in that attaches to any such off-the-shelf fixed-base model and gives it a mobile base, in zero shots, without any mobile manipulation expert data. The key idea is to treat the last-metre problem as aligning two keypoints: one on the target object and one on the robot arm, then solving for base and arm trajectories that bring them together while keeping the arm within a comfortable operating range. If this works, the large ecosystem of fixed-base manipulation models can be deployed for mobile tasks like 'serve me water' or 'prepare a meal' just by adding a navigation layer that knows where to dock. The paper reports higher success rates than existing mobile-manipulation systems on a standard benchmark and in real-world tests.","feed_headline":"Plug in MoTo, get a mobile robot arm — zero training","feed_subtitle":"It parks the base so the arm's keypoint reaches the object's, then any fixed-base model takes over.","key_machinery":"The central object is the interaction keypoint pair: a target keypoint (TK) on the object and an arm keypoint (AK) on the manipulator. MoTo aligns TK and AK through a constrained trajectory optimization that minimises their Euclidean distance subject to collision, smoothness, and arm-margin costs. The TK is produced by VLM keypoint proposal, keypoint selection, and multi-view 3D voting; the AK is obtained from wrist observation or end-effector proprioception. Aligning the two keypoints is what converts a navigation problem into a manipulation-feasible docking problem.","core_discovery":"MoTo (Move and Touch) is a plug-in module that takes any fixed-base manipulation model and turns it into a mobile manipulation system. The paper's central claim is that the problem of where to put the base can be solved zero-shot, without training on mobile manipulation demonstrations, by reducing every subtask to a keypoint-alignment problem. A VLM, prompted with an object-centric image and a language subtask, selects interaction keypoints on the target object and on the robot arm; multi-view voting fuses the 2D proposals into a 3D target keypoint. A trajectory optimizer then moves the base and arm so the arm keypoint reaches the object keypoint, while cost terms penalize collisions, sudden","pith_inferences":["Beyond the paper: the same keypoint-alignment reduction could apply to other mobile manipulation subproblems, such as tool-use over large surfaces or mobile assembly, as long as the subtask decomposes into reach-and-interact primitives.","Beyond the paper: the fixed-base reduction sets a testable boundary—tasks requiring simultaneous whole-body coordination, like opening a door with a rotating handle, will need MoTo combined with a whole-body control layer.","Beyond the paper: because MoTo depends on a pre-scanned static scene graph, an incremental online reconstruction module would extend it to dynamic environments; one could test whether refreshing keypoints at a higher frequency reduces failures from scene changes.","Beyond the paper: the multi-view voting threshold and the clustering resolution together determine keypoint precision; varying the threshold may reveal how sensitive success rate is to keypoint accuracy, a quantity the paper does not directly report."],"forward_implications":["Any fixed-base manipulation model—heuristic, diffusion-policy, or foundation model—can be lifted to mobile manipulation without retraining, so progress in fixed-base manipulation transfers directly to mobile tasks.","The approach does not require mobile manipulation expert data, lowering the cost of deploying mobile manipulation in new environments.","Docking points are chosen for interaction feasibility, not just proximity, addressing the compounding-error problem of naively chaining navigation and manipulation.","Multi-view voting can produce accurate 3D target keypoints from partial object images, enabling part-level interactions such as opening a fridge by its handle rather than its back.","Performance is bounded by the viewpoint generalization of the underlying fixed-base model; the paper finds 3D point-cloud-based policies generalize better across docking poses than some language-conditioned models."],"supporting_citations":[{"why":"Provides the mobile-manipulation benchmark, the baseline agent, and the four-stage evaluation used to measure MoTo's gains.","marker":"[18]"},{"why":"Introduces relational keypoint constraints for fine-grained manipulation, the direct inspiration for MoTo's interaction-keypoint framework.","marker":"[38]"},{"why":"An off-the-shelf vision-language-action manipulation model that MoTo plugs into and improves with its docking-point guidance.","marker":"[11]"},{"why":"A diffusion-based bimanual manipulation foundation model used as another plug-in backend in real-world tests.","marker":"[12]"},{"why":"Used to build the 3D scene point clouds and scene graphs that ground LLM task planning and keypoint generation.","marker":"[46]"},{"why":"Supplies the self-supervised pixel features that candidate keypoint proposals are clustered around.","marker":"[39]"},{"why":"Provides the part-level segmentation masks that constrain where keypoint proposals may lie.","marker":"[40]"},{"why":"The conventional visual-navigation baseline that MoTo replaces to show interaction-aware docking beats proximity-based navigation.","marker":"[44]"},{"why":"The inverse-kinematics solver that converts optimized joint trajectories into executable base and arm actions.","marker":"[41]"},{"why":"The Dual Annealing optimizer that searches the base and arm proposal space under the keypoint-alignment cost.","marker":"[42]"}],"fun_headline_variants":["MoTo turns any fixed-base arm into a mobile manipulator","Zero-shot mobile manipulation via keypoint alignment","No mobile training data? MoTo still gets your arm moving","Plug in MoTo, park the base, let the arm work"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"MoTo assumes every mobile task can be broken into sequential subtasks that each reduce to 'navigate to a spot, lock the base, then run a fixed-base manipulation,' and that all relevant objects are fully visible at sufficient resolution for keypoint extraction.","fun_headline_variants_meta":{"raw":{"variants":["MoTo turns any fixed-base arm into a mobile manipulator","Zero-shot mobile manipulation via keypoint alignment","No mobile training data? MoTo still gets your arm moving","Plug in MoTo, park the base, let the arm work"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000528,"raw_usage":{"total_tokens":2417,"prompt_tokens":812,"completion_tokens":1605,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1549}},"tokens_in":556,"tokens_out":1605,"duration_ms":16565,"temperature":1.0,"reasoning_tokens":1549,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:18:07.357654+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place the target object behind a partial occlusion so it is visible only at low resolution in each camera view; MoTo's keypoint proposal and voting should degrade sharply. If a single low-resolution view still yields successful docking and manipulation, the full-visibility assumption is not load-bearing.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used to build the 3D scene point clouds and scene graphs that ground LLM task planning and keypoint generation."},{"cited_title":"Kirillov, E","cited_arxiv_id":null,"evidence_quote":"Provides the part-level segmentation masks that constrain where keypoint proposals may lie."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The conventional visual-navigation baseline that MoTo replaces to show interaction-aware docking beats proximity-based navigation."},{"cited_title":"Carpentier, G","cited_arxiv_id":null,"evidence_quote":"The inverse-kinematics solver that converts optimized joint trajectories into executable base and arm actions."}],"review_version":1}