acceptodds
Under review as a conference paper at ICLR 2027

Beyond Arrival: Geometric Alignment for Agentic Humanoid Loco-manipulation

Abstract

Loco-manipulation for humanoid robots in open environments requires both long-range navigation and whole-body loco-manipulation abilities. While recent works have demonstrated powerful navigation and manipulation capabilities, bridging the two within a unified framework remains an open problem. The primary difficulty is that the navigation goal and the initial state of manipulation are misaligned in terms of observations. Existing approaches typically rely on task-specific training or additional annotations, with limited generalization across tasks and scenes. We introduce a handoff phase to reduce observation misalignment and use object-relative geometry as a bridge between navigation and manipulation. Building upon this insight, we propose GATE, a Geometry Aligned Task Execution bridge that leverages geometry aware representations to bridge navigation and manipulation, requiring no extra training or markers. An agent selects position and orientation references from the manipulation objective; a geometry aware feedback controller adjusts the robot's position before handoff. Our approach effectively improves the overall success rate of loco-manipulation tasks and demonstrates clear advantages over alternative methods in generalization, stability, and robustness. Building upon this approach, we develop an agentic system and deploy it in the real world. It completes multiple complex whole-body manipulation tasks with substantially higher success rates than the baselines. We further evaluate GATE on the OVMM benchmark, providing evidence that the geometric handoff principle is applicable to a different embodiment and an independently trained manipulation policy.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Humanoid Loco-manipulation, Agentic navigation, Mobile manipulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_XTM1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a "hand-off" phase for loco-manipulation task, as an transitioning stage between navigation and manipulation. This stage can help the robot to properly align the key objects in the comfortable positions (like front, center). This is done by aligning the key object poses and edges. The proposed agentic pipeline is evaluated with real robot and shows advantages over directly using a VLA.

Critical strengths

I believe the "hand-off" phase is a very good concept, practical for good loco-manip tasks.

Critical weaknesses

I generally do not find the right technical depth and sufficient novelty in this work for a venue like ICLR. I hope the "hand-off" phase can be more learnable in a data-driven manner unlike the proposed pipeline. This proposed system leans more towards an application like paper for IROS or ICRA.

Questions for authors

See weakness.

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

This paper proposes GATE, a geometry-based alignment module that bridges navigation and manipulation for humanoid robots. It uses task-conditioned object detection and rule-based geometric feedback to adjust the robot's position and orientation before manipulation. The authors integrate GATE into a dual-agent framework and demonstrate improved success rates on real-world humanoid tasks and the OVMM benchmark.

Critical strengths
  1. Practical relevance. The paper addresses an important problem in long-horizon loco-manipulation: the mismatch between navigation endpoints and the initial states required by manipulation policies.

  2. Good real-world results. The system demonstrates substantial improvements in task success rates on a real humanoid robot, supported by ablation studies and additional geometric analyses.

Critical weaknesses
  1. Limited novelty and technical insights. The paper primarily focuses on engineering integration of existing components, including VLM-based perception, object detection, rule-based geometric alignment, and pretrained control policies. While the system is effective, the proposed method is relatively straightforward and offers limited new algorithmic contributions or generalizable insights beyond existing navigation-to-manipulation approaches.

  2. Concerns about robustness and generalization. GATE relies heavily on accurate object detection, segmentation, and structural edge extraction, followed by hand-designed control rules and thresholds. Such a pipeline may be brittle under occlusion, detection errors, irregular object geometries, or changing environments. The current experiments do not sufficiently demonstrate robustness under these challenging conditions.

Questions for authors

Q1. What are the key technical contributions of GATE beyond existing geometry-based, manipulation-aware navigation approaches? What generalizable insights does the proposed method provide?

Q2. How robust is GATE to inaccurate object detection, missing or ambiguous structural edges, and irregularly shaped objects? Have the authors evaluated its sensitivity to perception errors and manually configured thresholds?

Reviewer_CtN4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the handoff between semantic navigation and manipulation in long-horizon embodied tasks, and proposes GATE, a training-free geometric alignment module. The module uses a vision agent, conditioned on the task, to select position and orientation references, and drives a base controller with RGB-D feedback to reduce forward-distance, lateral-offset, and yaw errors, with the goal of making navigation stopping points better suited to the subsequent manipulation. The final full system integrates the GATE module with navigation and manipulation policies. The paper conducts experiments on both a real robot and in simulation to verify the improvements brought by GATE. The real-robot experiments on the G1 humanoid verify the performance of the full system under the evaluated settings, and ablations support the effect of explicit alignment. The simulation experiments on the OVMM benchmark with the single-armed wheeled Stretch robot provide initial evidence that GATE can still improve performance under a different embodiment.

Critical strengths
  1. For the problem of stopping-point selection, the paper offers a solution that is sensible and effective under its experimental settings, providing a concrete interface between navigation and downstream whole-body skills. The real-robot ablation under the paper's setup shows that alignment does change the state at the manipulation handoff and brings measurable performance gains.
  2. The real-robot G1 evaluation and the simulated Stretch evaluation provide initial evidence for cross-embodiment adaptation. The videos in the supplementary material are impressive, especially the long-horizon real-robot tasks completed by the final full system, even though GATE participates only as one module within it.
Critical weaknesses

(1) The preconditions for the method to operate effectively are insufficiently validated The effectiveness of GATE rests on three preconditions: (i) the vision agent can select, within the current view, a visible position reference and an orientation reference carrying a visible straight structural edge (§3.3, Appendix B.1); (ii) the normal pointing toward that edge is a suitable desired heading for the base (Eqs. 11–13); and (iii) the preset standoff and position/orientation tolerances deliver the robot into the set of initial states acceptable to the downstream policy (Table 3). Once any of these preconditions fails, the method appears to degrade sharply. Taking assumption (i) as an example: in the existing experiments, the supporting surfaces of all four tasks have straight edges (Table 12; Figure 13); if a scene contains no visible straight-edge candidate (e.g., a round table, or only curved supporting surfaces), the orientation reference cannot be established, alignment cannot pass the readiness check, and the task can only fall back to navigation or fail. The ablations show that edge-based orientation alignment is precisely the main source of the gains (Table 15). The boundary between where this precondition holds and where it fails is left for the reader to infer from the current manuscript. The paper describes, at the mechanism level, how failures of these preconditions are handled, but it reports no statistics on alignment failures, retries, or NOT_FOUND frequency. As for the parameters, the 0.90 m standoff and the 0.10/0.08 m forward/lateral and 3° orientation tolerances are listed only as "current defaults" (Table 3); nowhere does the paper give the rationale for these values, nor any ablation or sensitivity study over them; the corresponding geometric parameters for the OVMM configuration are not disclosed at all. (2) The reported gains are difficult to attribute independently of the system configuration, external validity is hard to judge, and the statistical strength is limited GATE is a training-free module interposed between navigation and manipulation; its gains can be attributed only after decoupling from the upstream and downstream configurations. In the real-robot G1 experiments, however, both the upstream and the downstream are systems configured by the authors themselves, and the downstream is co-designed with GATE: the initial-state distribution on which the manipulation policy was trained may itself be correlated with GATE's output (C.7.1). This raises a hypothetical question: if the policy were trained on a broader initial-state distribution, or if navigation could stop more accurately at manipulable poses, would GATE's gains disappear? The paper in its current form cannot answer this question, which makes it difficult to judge the external validity of the method. The near-field experiments (Table 20) are a positive signal responding to this hypothetical concern, but they do not dispel my doubt, because they still use the task policy co-designed with GATE. Moreover, the paper only states that the evaluation region is "defined based on the typical distribution of navigation endpoints," without explaining how that distribution was determined; for reference, a typical ObjectNav configuration requires stopping within 1.0 m of the goal, with the goal visible from the stopping point (Habitat Challenge 2023, https://aihabitat.org/challenge/2023/), whereas the underlying navigation here declares arrival with a 2.0 m tolerance. At what stopping-accuracy level do the near-field initial positions sit? Would endpoints satisfying the community-standard accuracy still need GATE? The paper provides no examination of this. Furthermore, in the simulated OVMM experiments, after the upstream and downstream modules are replaced, GATE's Overall gain over the reproduced HomeRobot RL baseline shrinks to 3.17 percentage points (8.18 pp with sampling adaptation), markedly smaller than the improvement observed on G1. The OVMM evaluation uses a single seed; the authors' own results carry no standard errors, while the cited HELIOS results do, and the relation between the 3.17 pp gain and this noise level is not discussed. In addition, each task–method combination on the real robot has only 10 trials, without confidence intervals or significance tests, and the OVMM evaluation uses ground-truth semantic masks, the navigation mesh, and oracle picking (D.4). As it stands, the current evidence can hardly support the claims of generalization and robustness made in the abstract.

Questions for authors
  1. Current parameter settings: Table 3 lists the default stopping distance and tolerances for G1 but does not explain how the current values were determined; Appendices B.4/D.4 likewise do not give the corresponding geometric parameters used for OVMM. Could you explain the selection rule behind the G1 parameters, and report the actual stopping distance and the forward, lateral, and orientation tolerances used on OVMM? Was any sensitivity analysis conducted over different standoff or tolerance values?
  2. Near-field region representativeness: Appendix D.5 states that the near-field sampling region is defined by the typical distribution of navigation endpoints. Could you give the actual extent of this region and how it was determined? The description of demonstration starting headings in Appendix C.7.1 is qualitative; do the existing records support a quantitative comparison between the demonstration headings and the post-GATE aligned headings?
  3. Alignment reliability: The paper describes the recovery and fallback logic for when references are unavailable but does not report how often this actually occurs. Could the authors provide per-trial statistics broken down by task and method, including the counts of valid reference establishment, NOT_FOUND or edge-estimation failures, retries/recoveries, timeouts, readiness-check passes, and proceeding to manipulation despite alignment failure? If logs are readily available, please also report geometric-estimation errors or representative failure cases, so that we can judge how strongly the main results depend on these preconditions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.