acceptodds
Under review as a conference paper at ICLR 2027

Where to Stand, What to Grasp: Occupancy-Aware Coupling of Mobility and Manipulation

Abstract

Mobile manipulation is challenging not because navigation and manipulation are individually difficult, but because they are tightly coupled: the robot’s base pose determines whether the arm can reach the target. Yet prevailing vision-language-action (VLA) models largely collapse this coupling into a monolithic action head, treating robot pose as merely another concatenated feature. We present M-VLA, which makes three modifications to the standard trunk- plus-flow-expert architecture. First, a Grounded Geometric Tokenizer (GGT) enables geometry-aware representation by converting the SLAM occupancy map and robot poses into metric-scale-preserving tokens within a shared physical coordinate frame, allowing map geometry and robot localization to interact through joint attention in the backbone. Second, Localization-as-Operator Fusion represents robot localization not as a conventional feature but as an SE(2) transformation operator that transports intent between the navigation (world) and manipulation (body) frames. It also incorporates an uncertainty-derived trust scalar distilled from the otherwise discarded SLAM covariance. Third, Dual Functional Experts replace the single action head with separate navigation and manipulation flow-matching experts coupled by a Coordination Bridge, enabling each expert to pretrain on abundant single-domain data while exchanging task intent at every layer. On a large- scale, long-horizon mobile-manipulation benchmark comprising tasks lasting from one to more than fourteen minutes and requiring cross-room navigation across multiple indoor scenes, \texM-VLA} achieves the strongest overall performance.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot, VLA, Mobile manipulation

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_8RhjOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents M-VLA, a vision-language-action model for long-horizon mobile manipulation. The paper identifies the coupling between base placement and arm reachability as the central challenge: a navigation policy may bring the robot close to an object while still leaving the object unreachable or difficult to manipulate.

Critical strengths
  • The paper studies an important and practically relevant problem The coupling between navigation and manipulation is a meaningful problem in mobile robotics. In particular, the distinction between merely reaching a target region and reaching a pose that enables successful grasping or placement is important for long-horizon mobile manipulation. The paper provides a clear motivation for incorporating spatial maps, localization, and functional decomposition into a VLA policy.

  • The proposed system combines geometric observations with real-robot evaluation The use of SLAM occupancy maps and localization uncertainty is intuitively appropriate for this setting. The separation of navigation and manipulation experts is also a reasonable architectural direction. In addition, the paper includes real-robot experiments instead of evaluating only in simulation, which makes the proposed system potentially relevant to practical deployment.

Critical weaknesses
  1. The central SE(2)SE(2) operator contribution is not clearly operationalized. The main technical concern is that the claimed Localization-as-Operator Fusion is not sufficiently implemented or specified in the forward computation. Equation (1) defines a coordinate transformation, pB=Rt⊤(pW−pt)p^B=R_t^\top(p^W-p_t), but the paper never explains where the world-frame intent pWp^W is generated, where the transformed body-frame quantity pBp^B is consumed, or how this transformation is integrated into the interaction between the navigation and manipulation experts. In the actual implementation described by Eqs. (3)–(5), the localization information appears to be represented as a raw pose, a trust scalar, and a validity bit, which are then processed by Fourier features and an MLP to form a prefix token. This is closer to pose-conditioned feature injection than to an explicitly applied SE(2)SE(2) operator. The paper does not provide a transformed waypoint, a body-frame manipulation constraint, an equivariant attention operation, or a frame-consistent training objective. Therefore, the claims about explicit intent transport and position invariance are not established by the presented method. The M-VLA-concat comparison also cannot isolate the effect of an explicit operator, since the two variants may differ in tokenization, Fourier encoding, trust information, and localization injection rather than only in the use of an SE(2)SE(2) transformation.

  2. The related-work discussion is incomplete and does not sufficiently cover directly relevant mobile-manipulation literature. The paper cites a large number of broadly related VLA, navigation, locomotion, and foundation-model works, but it omits several papers that are more directly connected to its stated problem of manipulation-aware base placement and mobile-manipulation policy learning. In particular, the authors should discuss MoMa-Kitchen [1], which studies affordance-grounded last-mile navigation and the selection of final robot positions that enable subsequent manipulation; MoManipVLA [2], which transfers fixed-base VLA capabilities to mobile manipulation through coordinated base and arm trajectory optimization; Mobile ALOHA [3], which is a representative system for bimanual mobile manipulation and whole-body teleoperation; and HomeRobot [4], which establishes a standard open-vocabulary mobile-manipulation setting involving navigation, object interaction, and placement. These works are directly relevant to the paper’s motivation, data collection protocol, benchmark positioning, and novelty claims. The omission is especially important because the paper introduces an affordance field and repeatedly argues that the navigation policy should select manipulation-feasible base positions, yet it does not discuss closely related affordance-grounded positioning work. The related-work section should explain how M-VLA differs from affordance-based final-position prediction, explicit base-placement optimization, fixed-base VLA transfer, whole-body teleoperation, and open-vocabulary mobile-manipulation benchmarks. The relevance of these omitted works is reflected in their direct focus on mobile manipulation, final base positioning, and VLA-based policy transfer.

[1] MoMa-Kitchen: A 100K+ Benchmark for Affordance-Grounded Last-Mile Navigation in Mobile Manipulation ICCV2025

[2] MoManipVLA: Transferring Vision-language-action Models for General Mobile Manipulation CVPR2025

[3] Mobile ALOHA Learning Bimanual Mobile Manipulation with Low-Cost Whole-Body Teleoperation CoRL2024

[4] HomeRobot: Open-Vocabulary Mobile Manipulation CoRL2023

  1. The contribution statement is overly component-centric and does not make the actual novelty clear. The paper is written in a component-by-component manner that reads more like an engineering inventory than a coherent scientific argument. The five contribution bullets largely restate the modules introduced in the method section: M-VLA itself, GGT, Localization-as-Operator Fusion, Dual Functional Experts, and the final empirical results. However, the paper does not clearly articulate the underlying scientific insight, the hypothesis being tested, or the specific conceptual advance over prior work.

  2. The baseline comparison and ablation studies do not support clear causal attribution. The paper does not clearly specify whether π0.5\pi_{0.5}, Task-VLA, and StarVLA-π\pi receive the same SLAM occupancy maps and localization inputs as M-VLA, nor does it provide sufficient information about their training data, optimization schedules, model sizes, or hyperparameter tuning. If the baselines use only RGB and proprioception while M-VLA additionally receives occupancy maps and pose estimates, the comparison mixes the effect of the proposed architecture with the benefit of additional geometric modalities. The ablations have similar confounding factors. No GGT removes occupancy and geometric tokens jointly, No Fusion removes several localization-related signals at once, and No Dual Experts changes the action-head structure and potentially the model capacity. The paper does not separately evaluate the trust scalar, pose encoding, occupancy map, coordination bridge, or explicit frame transformation. It also lacks a parameter-matched single-expert baseline and a dual-expert model without cross-expert interaction. As a result, the reported ablations show that the complete system is useful, but they do not demonstrate which individual design choice is responsible for the improvement.

Questions for authors

See weakness.

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

This paper studies base–arm coordination in mobile manipulation, emphasizing that base placement determines the feasibility of subsequent manipulation. M-VLA augments a pretrained vision-language backbone with occupancy-map and robot-state tokens that encode metric spatial information, represents planar pose and covariance-derived localization trust through a shared token, and introduces navigation and manipulation flow-matching experts that interact through attention. The training scheme supports single-domain pretraining followed by joint training on mobile-manipulation demonstrations. The authors evaluate the method on long-horizon BEHAVIOR-1K tasks and two real-robot task families, reporting improvements over the evaluated VLA baselines. Component ablations and a concatenation variant using the same SLAM inputs provide initial support for the approach.

Critical strengths

S1. The paper addresses an important problem with an architecture reasonably aligned with the task requirements. Since base placement affects manipulation success, allowing navigation and manipulation representations to interact while providing metric maps and localization reliability is well motivated. The functional expert split also provides a viable way to use single-domain demonstrations, reducing the need for every training example to contain complete mobile-manipulation behavior.

S2. The method is evaluated in both long-horizon simulation and real-robot experiments, with some component-level evidence. The paper reports an average living-space success rate of 0.32, compared with 0.22 for the strongest evaluated baseline. With 30 real-robot trials per task per method, success increases from 0.57 to 0.70 for Package Sorting and from 0.53 to 0.63 for Shelf Placement.

Critical weaknesses

W1. Comparison with staged mobile-manipulation approaches is missing. The paper emphasizes the advantages of joint navigation–manipulation decision-making, but does not compare against staged approaches such as N2M [1], which select an initial manipulation pose before executing manipulation. The advantage of the joint architecture over this class of approaches therefore remains insufficiently established.

[1] Kaixin Chai, Hyunjun Lee, and Joseph J. Lim. N2M: Bridging Navigation and Manipulation by Learning Pose Preference from Rollout. ICML, 2026.

W2. The experiments do not sufficiently validate the base–arm coordination mechanism. The existing ablations and the concatenation comparison using the same SLAM inputs show overall benefits, but lack a controlled experiment that retains both experts while disabling only the Bridge, as well as a direct analysis of how base placement affects manipulation success.

Questions for authors
  1. Comparison with staged approaches: Could you provide a comparison with placement-then-manipulation approaches such as N2M to validate the advantages of joint decision-making?

  2. Validation of the coordination mechanism: Could you provide an ablation that retains both experts and all geometric inputs while disabling only the Coordination Bridge, and analyze how base placement affects subsequent manipulation success?

  3. Implementation clarification: Which tensors does the coordinate transformation in Eq. (1) act on, and how are the transformed results used by the Coordination Bridge?

  4. Figure notation and clarity: Should the “SE(3)” label in Figure 1 be changed to “SE(2),” as used in the main text and Appendix A.4? In addition, some figures have low visual clarity, making small text and annotations difficult to read. Please improve their resolution or use vector graphics.

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

The paper proposes M-VLA, a vision-language-action model for mobile manipulation that treats base–arm coupling as a first-class architectural construct, mainly including a Grounded Geometric Tokenizer (GGT), Localization-as-Operator Fusion and Dual Functional Experts. Specifically, GGT allows map geometry and robot localization to interact through joint attention in the backbone; Localization-as-Operator Fusion uses robot localization as an SE(2) transformation to transfer intent between navigation and manipulation frames, together with an uncertainty-derived trust scalar; Dual Functional Experts enables each expert to pretrain on abundant single-domain data while exchanging task intent at every layer. Experimental results on the BEHAVIOR-1K environment (simulated in Omni-Gibson) and a real-world robot validate the performance of the method.

Critical strengths
  1. Treating base–arm coupling as a first-class architectural construct is an interesting idea.
  2. The method shows clear performance gains over baselines.
Critical weaknesses
  1. The core technologies proposed in the paper lack detailed explanations. One of the contributions in the paper is to promote the base–arm coupling to a first-class architectural object rather than absorbing it into a single action head. The proposed three modifications include a Grounded Geometric Tokenizer (GGT), Localization-as-Operator Fusion and Dual Functional Experts. However, navigation and manipulation are two different tasks, and their inputs and outputs are always different. How are the information from different models integrated together by the three proposed modules? It is unclear according to the current expressions. For each module, it lacks some details and may be difficult for understanding and algorithm reproduction. For instance, the paper stated ‘The robot pose (x, y, yaw) is encoded directly into the embedding space. Each continuous quantity is encoded with a multi-band Fourier encoding; the results are concatenated and passed through MLPs to produce geometric tokens.’ How can these be achieved? This is just an example and there are still some other unclear expressions in the paper.
  2. Some symbols in the equations of paper are not defined, which hinders the understanding of the core algorithm of the paper. Such as Rt and Pt in Eq (1), W1 and W2 in Eq (4), Lefm in Eq (6), and so on. These equations represent the core algorithm. The current presentation and explanation make it difficult for readers to understand the specific details of the algorithm.
Questions for authors
  1. What is the meaning of “a first-class architectural object”? What is the connection between it and “base–arm coupling”?
  2. The paper stated ‘the resolution and origin position are no longer discarded’. How are these reflected in the algorithm or the corresponding equations?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.