acceptodds
Under review as a conference paper at ICLR 2027

KI-VLA: Kinematics-Informed Action Decoding for Flow-Matching Robot Policies

Abstract

Flow-matching robot policies can generate operational-space action chunks without explicitly modeling the underlying joint-space motion required for execution. As a result, decoded chunks can contain jitter, micro-reversals, and locally inconsistent commands that degrade closed-loop performance. We introduce KI-VLA (Kinematics-Informed VLA), a predictor-corrector decoder that treats the flow policy's intermediate end-effector action chunks as noisy kinematic observations and uses them to infer a smooth latent joint-increment trajectory. A kinematic observation model and Kalman/RTS smoothing infer temporally coherent joint motion, which is mapped back to action space via bounded trust-region updates. By interleaving these corrections with flow-matching steps, KI-VLA refines intermediate proposals while keeping the base policy frozen. On RoboTwin 2.0, KI-VLA improves the baseline from to in the benchmark's Easy setting and from to in its Hard setting. On real hardware, clean-condition success is comparable ( to ), while success under a controlled disturbance injected into the action proposal rises from to . Further studies show that the gains are concentrated in the few-step decoding regime and can diminish under severe measurement-covariance or process-noise mis-scaling, clarifying the operating regime in which decoding-time kinematic correction is most useful.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, robot learning, imitation learning, flow matching, robot manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RkF1OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes Kinematics-informed VLA,which is a flow-matching robot policy. it treats the policy's end effector action chance as a noisy kinematic observation of an underlying joint-increment trajectory. The paper has several contributions:

  1. The paper uses Jacobian observation model and differentiable Kalman/RTS inference to estimate smooth joint-increment trajectories, then map them back to action space with bounded corrections.
  2. It interleaves correction with flow-matching updates and support sparse schedules that trade computation for performance, while keeping the base policy frozen.
  3. the paper evaluate the controlled end-pose VLA comparison on eight RoboTwin 2.0 tasks and four real-robot tasks, measuring task success and command-level motion.
Critical strengths

The research problem is interesting. Though I have seen some same topic papers already in this field. The integration of joint space kinematic consistency into intermediate flow matching is kind of novel.

Critical weaknesses

What's the intuition to insert Kalman smoothing between flow machine steps. The generalizability needs further investigation and demonstration. Are you training separate models for different tasks? or train a single model for different tasks?

Questions for authors
  1. Have you ever tried pi_0.5 as baseline? I think pi_0.5 is considered as a stronger model than pi_0. but table 1 only compare pi_0. I am wondering if Pi_0.5 this paper can beat the pi_0.5 model.

  2. Evaluating the method on another flow-matching policy could help demonstrate its generalizability beyond pi_0 backbone.

  3. It may worth testing on diffusion based policy

  4. Looks like this model takes both language instruction and RGB image as input, are they both input of the frozen flow policy module. I am wondering if it working well without a language encoder?

Reviewer_D9QxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes KI-VLA, a decoding-time correction method for flow-matching VLA policies. It treats intermediate end-effector action predictions as noisy observations of joint movements, uses a learned Jacobian with Kalman filtering and RTS smoothing to infer smoother joint trajectories, and maps them back to Cartesian actions with bounded trust-region corrections. This process is repeated during flow-matching sampling without retraining the base policy.

The method is tested on eight RoboTwin 2.0 tasks and four real-robot tasks, showing improvements in simulation and stronger gains under artificially injected action disturbances, but only modest improvements in clean real-robot settings.

Critical strengths
  1. Integration of classical robotics methods. The paper brings Jacobian-based kinematics and Kalman/RTS smoothing into flow-matching action decoding. While the individual components are well established, combining them with generative policies is an interesting direction.

  2. No need to retrain the base policy. The method keeps the pretrained VLA frozen and introduces relatively lightweight auxiliary modules. This could be useful when retraining a large policy is expensive or impractical.

  3. Consistent improvements in simulation. KI-VLA improves average success rates on both Easy and Hard RoboTwin settings, with improvements across most evaluated tasks. These results suggest that inference-time trajectory correction can offer some practical benefits.

  4. Detailed ablations and sensitivity analysis. The paper includes comparisons with conventional smoothing methods, internal component ablations, correction schedules, and different numbers of flow steps. These experiments help clarify where the method works and where its benefits are limited.

Critical weaknesses
  1. Unclear problem motivation. The paper attributes action jitter and discontinuities to the lack of explicit joint-space modeling, but provides little evidence to support this connection. It is unclear whether these issues arise from kinematic inconsistency or simply action-prediction noise, and how much they actually contribute to task failures.

  2. Lack of justification for the action representation. Many existing robot policies already predict joint-space actions, and temporal consistency can also be encouraged through action representations or training objectives. The paper does not sufficiently explain why a separate inference-time correction module is needed.

  3. Questionable role of the learned Jacobian. The Jacobian describes robot kinematics, not the noise introduced during flow matching. It is unclear why applying a Jacobian trained on demonstration trajectories to partially noisy intermediate samples is the right approach. Moreover, the analytic Jacobian achieves nearly identical results (50/35 vs. 51/35), making the benefit of learning it questionable. If action consistency is the goal, why not directly learn a transition or consistency model?

  4. Insufficient motivation for per-step correction. Intermediate flow-matching samples are noisy by design and do not necessarily represent meaningful robot trajectories. The paper does not convincingly explain why enforcing kinematic and temporal consistency at every denoising step is necessary or beneficial.

  5. Inconclusive correction-schedule ablation. Comparing a single final correction with ten intermediate corrections changes both the correction timing and frequency. This makes it difficult to determine whether the improvement actually comes from correcting intermediate samples.

  6. Unnecessary methodological complexity. The method combines a learned Jacobian, uncertainty estimation, Kalman/RTS smoothing, relinearization, and trust-region projection for what is essentially a trajectory-smoothing problem. Stronger comparisons with simpler joint-space optimization or training-time regularization are needed to justify this complexity.

  7. Limited experimental evidence. The clean real-robot improvement is not statistically significant, while the larger robustness gains rely on artificially corrupted action proposals. The improvement also disappears at 50 flow steps on Turn Switch, raising the question of whether the method mainly compensates for low-step sampling artifacts rather than addressing a fundamental limitation of VLA policies.

Questions for authors
  1. Problem motivation. What evidence shows that action jitter and discontinuities come specifically from kinematic inconsistency rather than general prediction noise? Can the authors demonstrate how these issues lead to actual task failures?

  2. Simpler alternatives and learned Jacobian. Why is a learned Jacobian necessary when the analytic one achieves nearly identical results? How does the method compare with simpler alternatives such as joint-space action prediction, learned transition models, or training-time temporal regularization?

  3. Intermediate flow correction. What justifies treating partially denoised flow samples as meaningful kinematic observations? Why should temporal consistency be enforced throughout generation rather than only on the final predicted actions?

  4. Correction schedule and complexity. Could the authors compare correction at different sampling stages with matched correction counts and compute budgets? What does the full Kalman/RTS pipeline offer beyond simpler joint-space trajectory optimization?

  5. Experimental validation. Given the insignificant clean real-robot gains and the reliance on synthetic proposal disturbances, how effective is the method under normal execution? Why does its advantage disappear at 50 flow steps on Turn Switch, and could the gains mainly come from correcting low-step sampling artifacts?

1 comment

Thanks for the detailed responses and additional experiments. I appreciate the new results with π0.5 and the comparisons using the same number of correction steps. These address some of my questions and provide stronger evidence that the method can help. However, I’m still not convinced that the full pipeline is sufficiently justified.

My main concern is how much the kinematic component actually adds. The same interleaved smoother in action space gets 48/31, compared with 51/35 for the full method—a difference the response acknowledges is within noise. The analytic Jacobian also performs almost identically to the learned one. Given that these ablations are mainly on one task, I still find it hard to tell which parts of the pipeline are necessary and whether the added complexity brings a reliable benefit over simpler smoothing.

I’m also still concerned about how intermediate flow samples are treated. The response acknowledges that these drafts have a progress-dependent shrinkage that the observation model does not account for. Giving early drafts less weight does not resolve that mismatch. The new experiments suggest that interleaving corrections helps on Turn Switch, but they do not fully justify this particular observation model. Correcting the predicted endpoint seems like a relevant alternative, but it remains untested.

The practical benefit under normal execution also remains unclear to me. The simulation gains across two backbones are encouraging, but clean real-robot success only changes from 68.5% to 71.0%, without a statistically significant improvement. The larger gains under an injected disturbance are useful evidence, but they do not show how often this issue causes failures in normal execution. The gain also disappears at 50 flow steps on Turn Switch, which leaves questions about how broadly the method is useful.

Finally, I still find the paper difficult to follow. The rebuttal makes several points clearer, but the paper itself needs to explain more directly why each component is needed and which results support it.

Overall, I appreciate the authors’ efforts, but my main concerns remain unresolved, so I will keep my original rating

Reviewer_gbc1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper focuses on action-smoothing problems in robotic policy learning by proposing KI-VLA, a kinematics-informed predictor-corrector framework for flow-matching VLA policies that reduces action jitter and inconsistent robot movements without retraining the base policy. It learns a Jacobian model that connects joint-space motion to end-effector movement, then applies uncertainty-aware Kalman filtering and RTS smoothing to produce more temporally consistent trajectories. These trajectories guide bounded corrections interleaved with flow-matching decoding. Experiments show improved success rates on RoboTwin 2.0 and under real-world action disturbances, with the largest benefits in few-step decoding.

Critical strengths

I found the following strengths:

  • The framework effectively leverages a learned Jacobian and uncertainty-aware Kalman/RTS smoothing to infer coherent joint-space motion. This intuitively offers a more robust alternative to purely Cartesian action smoothing.

  • The method also regularizes intermediate action proposals during the flow-matching decoding process, which is bounded by a trust region, which seems to help refine trajectories before final execution.

  • The experiments cover soeveral comparisons across eight RoboTwin 2.0 tasks and several real-world hardware tests that successfully demonstrate the system's ability to reject injected action disturbances.

Critical weaknesses

The paper still has the following key isssues;

  • First, the method's advantages disappear at higher sampling steps (e.g., N=50N=50) and are highly sensitive to the calibration of measurement and process covariances. A discussion on this effect is required.

  • Insufficient Baselines: The evaluation omits critical comparisons against relevant test-time refinement methods (e.g., SUREFlow, LP-DS), post-optimization frameworks like LiPo, and a faithful implementation of ACG.

  • Confounding Variables in Ablations: The comparison between interleaved and final-only correction are not enough to separate the timing of the correction from the total number of corrector calls applied.

  • Also, the paper uses a constant-velocity kinematic prior without modeling dynamics or contact constraints, which can erroneously suppress necessary high-frequency micro-motions, such as rapid taps.

Questions for authors

I would revise my score if the following questions are well addressed:

  • How sensitive is the method's performance to the choice and scaling of process noise (QQ), and can you provide an automated calibration procedure?

  • In environments where an analytic Jacobian is available, are there proven benefits to learning a surrogate Jacobian (such as capturing backlash), or is it strictly for convenience?

  • Can the corrector detect and preserve legitimately needed high-frequency motions by adapting its uncertainty parameters online, or is a task-specific toggle required?

  • Did you observe numerical conditioning issues in the RTS smoother near singular configurations, and could alternative smoothers (like MBF) provide better numerical robustness?

  • And finally, can author update the simulation results on RoboTwin 2.0 with few more advanced method such as SUREFlow or LP-DS?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.