acceptodds
Under review as a conference paper at ICLR 2027

Action-Manifold VLA: Geometry-Aware Flow Matching in Joint and Task Spaces

Abstract

Diffusion- and flow-based action experts have become a dominant paradigm for vision-language-action (VLA) models, yet they hinge on two design choices: what geometry the action representation should obey, since a bimanual action chunk naturally lives on a product manifold rather than in Euclidean space, and in which space the generative dynamics should evolve, where joint space and task space each offer trade-offs. To study these questions, we propose Action-Manifold VLA (AM-VLA), a framework with two variants that share the same backbone and a unified action space but differ in where the flow lives. AM-VLA-J builds a joint-space flow on joint residual and gripper channels while introducing manifold-aware task-space auxiliary supervision, whereas AM-VLA-E builds a task-space flow on the per-arm end-effector residual manifold , modeling gripper, position residual, and rotation residual; rotations use shortest-geodesic conditional paths and Log-map target fields with time-constant norm. On the 50-task RoboTwin 2.0 benchmark, AM-VLA-J and AM-VLA-E achieve average success rates of and , respectively, compared with for the backbone, showing that manifold-aware task-space supervision improves joint-space generation, while manifold-native dynamics provide substantially larger gains when the generative process evolves directly in task space.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, Manifold, Flow matching

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_sxPwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper investigates the relationship between the generative space and the geometry of actions within Vision-Language-Action (VLA) models. Building on HoloBrain-0, the authors propose Action-Manifold VLA (AM-VLA), which reformulates diffusion-based action generation as conditional flow matching, and instantiate two variants: AM-VLA-J and AM-VLA-E. Experiments in simulation and on real robots show modest performance gains over HoloBrain-0.

Critical strengths

The paper features a clear structural design that explicitly distinguishes between generation states and conditional information. By developing generation schemes for both joint space and task space within a shared architecture, it offers a valuable research perspective for understanding the relationship between the generation space and the geometry of actions.

Critical weaknesses
  1. The main concern is that the advance over existing work is not yet sufficiently clear. RFMP/SRFMP[1] already performs manifold action generation on product spaces containing position, rotation, and gripper variables, but is not discussed. HoloBrain-0 also includes joint/pose prediction and task-space supervision. Distinguishing inherited designs from new ones, and identifying the experiments supporting the latter, would clarify the contribution. While the provided ablation studies demonstrate the utility of these components, they do not adequately prove that the new designs are superior to the existing designs already present in the base model.
  2. The performance gains are limited: showing improvements of only 0.4 and 1.0 percentage points on full benchmarks—and no significant benefit is observed in real-world grasping tasks. Are these gains consistently beyond random fluctuation, and can they be replicated across a broader range of tasks?
  3. The differences between the two generation schemes in terms of computational cost and control efficiency are not discussed, nor is it addressed whether the marginal performance gains are sufficient to justify the additional costs.
  4. The AM-VLA-E relies on a task-space controller; however, its controller or inverse kinematics (IK) implementation, handling of unreachable targets, and behavior in the event of execution failure are not sufficiently described.

[1] Fast and Robust Visuomotor Riemannian Flow Matching Policy arXiv:2412.10855

Questions for authors
  1. What is the central contribution beyond HoloBrain-0 and RFMP/SRFMP, and which experiment independently supports it?
  2. How do AM-VLA-J, AM-VLA-E, and HoloBrain-0 compare in inference latency, and effective control frequency?
  3. What controller or IK solver is used to execute AM-VLA-E outputs? How are unreachable poses, joint-limit violations, IK failures, and execution failures handled?
  4. The long-horizon benefit claimed in Section 4.2 is not reliably established, as it is supported mainly by a single stacking task. Could the authors validate this claim on additional multi-stage tasks (Sim or Real world)?
1 comment

I think the author has given up on rebuttal; I will keep my current rating.

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

This paper proposes Action-Manifold VLA (AM-VLA) to study how generation space and action geometry affect VLA policies. AM-VLA-J performs joint-space flow matching with auxiliary pose and differentiable forward kinematics (FK) supervision. AM-VLA-E generates end-effector residuals using SO(3) geodesic paths and Lie-group updates. On RoboTwin 2.0, they achieve average success rates of 92.5% and 93.1%, compared with 92.1% for HoloBrain-0.

The difficult-task ablations are encouraging, but the contribution beyond existing manifold policies and the implementation of the key rotational baseline need clarification.

Critical strengths

S1. The paper addresses a practical problem with a clear design.

Table 1 and the method section distinguish generated variables, conditioning variables, and executed actions. The different roles of auxiliary pose prediction and differentiable FK supervision are clearly explained.

S2. The difficult-task ablations provide positive evidence.

In Table 4, the SO(3) variant improves average success from 69.4% to 81.6% over the Euclidean quaternion variant, with gains on all five tasks. Table 3 also supports the benefit of FK supervision. The task-selection rule is explicitly stated.

S3. The proposed method is mathematically well specified and includes real-robot validation.

The appendix defines the proposed method's quaternion conventions, prediction-to-velocity conversion, and training and sampling procedures. The real-robot experiments further demonstrate deployability, with AM-VLA-E achieving the highest observed success rate on both tasks.

Critical weaknesses

W1. The distinction from existing Riemannian flow matching policies needs clarification.

The paper acknowledges Riemannian Flow Matching but omits directly relevant work, including Braun et al., Riemannian Flow Matching Policy for Robot Motion Learning, and Ding et al., Fast and Robust Visuomotor Riemannian Flow Matching Policy. These methods already apply Riemannian flow matching to robot policies and end-effector pose generation.

Please clarify which technical designs and empirical insights in VLA adaptation, residual action modeling, and joint versus task-space generation extend beyond this work.

W2. The Euclidean quaternion baseline in the key ablation is insufficiently specified.

The 12.2-point gain in Table 4 is central to the paper's conclusions. The paper keeps the task-space generation state and execution interface unchanged, and Appendix B.2 provides general quaternion conventions. However, their application to the Euclidean baseline remains unclear.

Please specify its rotational prior, normalization during the path, loss scaling, and sampling step count, alongside the SO(3) configuration. These details are needed to assess which design changes account for the gain.

W3. The presentation could be more polished.

Some figures are visually crowded, with small annotations, and parts of the method description are repetitive. More concise writing and clearer figure layouts would help highlight the main designs and findings.

Questions for authors

Q1. What technical designs and empirical findings are new relative to RFMP?

Q2. How do the Euclidean and SO(3) variants differ in their priors, sign handling, normalization, loss weights, and sampling steps?

Q3. Do Tables 3–4 use five-task retraining or subset evaluation of the 50-task models, and which evaluation conditions are used? Please report training repetitions and uncertainty for the key performance differences.

Reviewer_d6DKOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the choice of action generation space and how to account for the geometric structure of actions during generation. Building on HoloBrain-0, the paper proposes two variants: AM-VLA-J, which predicts joint-space residuals, and AM-VLA-E, which predicts end-effector pose residuals. AM-VLA-E further adopts an SO(3) geodesic formulation for rotation prediction. Through experiments and ablation studies on RoboTwin and real-world robots, the paper demonstrates the feasibility of the proposed methods.

Critical strengths
  1. The simulation and real-world experiments demonstrate the feasibility of using flow matching for action generation within the HoloBrain-0 framework.
  2. The ablation studies on AM-VLA-E demonstrate the benefits of using SO(3) geodesic for rotation prediction.
Critical weaknesses
  1. Most core components, including residual actions and FK-based conditioning and supervision, are inherited from HoloBrain-0. AM-VLA-J primarily reformulates the action expert using standard flow matching, while Table 3 mainly ablates pose and FK supervision already present in HoloBrain-0. These results demonstrate the effect of inherited components rather than the benefits of the proposed modifications.
  2. The paper proposes an “action-manifold” framework, but the joint-space/task-space distinction is essentially a choice between predicting joint states and end-effector poses. Meanwhile, the rotation modeling in AM-VLA-E also follows existing SO(3) flow-matching techniques.
Questions for authors
  1. Which mechanisms are newly introduced beyond HoloBrain-0? Are components such as FK supervision and auxiliary pose supervision unchanged from HoloBrain-0, or have they been modified? If modified, what are the specific changes?
  2. Given the small improvement of AM-VLA-J over HoloBrain-0, how does AM-VLA-E compare with a HoloBrain-0 variant that directly generates end-effector poses while retaining its original diffusion formulation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.