acceptodds
Under review as a conference paper at ICLR 2027

GFC: Geometry-First Corruption for Generative Imitation Learning

Abstract

Generative policies model expressive action distributions for robotic manipulation, but standard Gaussian corruption does not explicitly account for dependencies across the action horizon. To address this issue, we propose GFC: geometry-first corruption, which generates noisy trajectories by randomly deforming a trajectory-level Riemannian metric. The metric captures action geometry and temporal coupling, refined by cross-segment predictive responses to structured perturbations. At each forward step, a random self-adjoint field deforms the metric through a matrix exponential, and a metric-gradient flow induces trajectory motion without an additional coordinate-noise draw. This connects demonstration geometry along the action horizon with stochastic corruption along noising time. We derive path-wise bounds on trajectory motion and predictive response, and specify reverse-learning objectives and terminal-source requirements. Experiments in simulation and on real robots demonstrate improved trajectory quality and policy performance. More broadly, GFC establishes stochastic geometry as a design principle for generative modeling, allowing learned data structure to shape the corruption process itself. Videos and supplementary materials are available at https://anonymous.4open.science/r/GFC/.

Primary area:
generative models
Keywords:
Riemannian Manifold, Geometry-aware Diffusion, Imitation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_Z3KjOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

GFC introduces a geometry-aware corruption process for generative imitation learning, using learned trajectory dependencies and random metric deformations to produce structured perturbations for reverse policy learning.

Critical strengths

The controlled comparisons are valuable and the evaluation is comrehensive.

Critical weaknesses
  1. The relationship between response learning, metric refinement, random deformation, source fitting, and reverse generation should be explained more intuitively.

  2. The practical specification of the terminal source remains insufficiently clear. The paper notes that the proposed forward process does not automatically converge to a known Gaussian, but the conditional source model used for robot experiments requires further implementation details and validation.

  3. The real-world Diffusion Policy baseline performs unexpectedly poorly, achieving only 2/100 successes on T-block pushing and 0/100 on box uprighting. Further evidence that the baseline was adequately trained and evaluated is important for interpreting the large performance gap.

  4. Some experimental descriptions are inconsistent. For example, Section 4.4 reports 200 demonstrations per real-world task, whereas Appendix E.6 reports 100 demonstrations for box uprighting.

Questions for authors

please address the concerns above

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

This paper proposes Geometry-First Corruption (GFC), a forward corruption process for diffusion and flow-based imitation policies. Instead of adding Gaussian noise, GFC randomly deforms a Riemannian metric over action trajectories and corrupts trajectories by following a gradient flow under this deformed metric. The method is evaluated on synthetic trajectories, simulated manipulation and navigation benchmarks, and two real-robot tasks.

Critical strengths
  • The idea of shaping the forward corruption through a random trajectory metric is interesting.
  • The evaluation is broad and includes real-robot experiments.
Critical weaknesses
  • The paper is very hard to follow. The method section is extremely dense and gives almost no intuition about what the corruption does or why it is preferable to structured noise. The writing must be significantly improved, e.g. with an intuitive overview and a pipeline figure, so that the reader can follow the proposed changes.
  • The notation is confusing and inconsistent. Eq. 1 uses symbols before they are defined, ψ\psi and ω\omega are overloaded, Eq. 3 uses Δt\Delta t while the text defines Δh\Delta_h, and much of the required notation is only introduced in Appendix B.
  • I did not review the mathematical derivations thoroughly, and I do not have a deep understanding of this field. My assessment is mainly based on the presentation and the experiments. However, I note that the poor presentation and unintuitive writing worsen this, to the extent that I had to spend a significant amount of time on just interpreting the main arguments.
  • The experimental evaluation also has several weaknesses.
    • GFC’s terminal distribution is not Gaussian and requires a separately fitted terminal source, while the baselines sample from a standard normal. Is the comparison then still controlled, as claimed in App. E.4?
    • The number of seeds is never stated, and Tables 2 and 4 report no variance.
    • On the real robot, Diffusion Policy succeeds in only 2/100 and 0/100 trials, although it solved real-world Push-T at 95% in its original paper. Was this baseline tuned?
    • There are inconsistencies, e.g. the number of demonstrations differs between Sec. 4.4 and App. E.6, and the SO(3) results for two different metrics in Table 1 are identical.
Questions for authors

Please see weaknesses.

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

The paper proposes Geometry-First Corruption (GFC), a forward corruption process for robot imitation learning that warps a Riemannian metric rather than adding independent Gaussian noise directly to coordinates. The metric penalizes abrupt changes between consecutive actions, as well as perturbations that disrupt the dependencies between trajectory segments. GFC then creates corrupted trajectories via metric-gradient flow on an energy landscape defined by the deformed Riemannian metric. Experiments show GFC achieves lower local-geometry distortion, roughness, and path-length distortion on \mathbb{R}^2, S^2, and SO(3) spaces than coordinate-based corruption. On learned reverse generation, it has the lowest recovery RMS error and local-geometry distortion on \mathbb{R}^2 and SO(3), and yields the lowest SWD and MMD on \mathbb{R}^2. On robot imitation tasks, GFC also shows lower jerk and higher scores and success rates than other forward corruption methods on most tasks. It also demonstrates a better success rate than Diffusion Policy in real-world experiments, including T-block pushing and box uprighting.

Critical strengths

The paper proposes an innovative forward corruption method that deforms a Riemannian metric, which penalizes abrupt changes between consecutive actions and perturbations that disrupt the dependencies between trajectory segments, instead of directly adding Gaussian noise to trajectories. Corrupted trajectories used for training are obtained by gradient flow on the landscape defined by the deformed Riemannian metric. This method embeds the structure of trajectories in the geometry of the space itself, providing a novel way of incorporating such structure into trajectory generation.

The experiments compare different forward corruption methods on various spaces and robot tasks. GFC outperforms Diffusion Policy, Flow Matching Policy, and Streaming Flow Policy on Push-T, and outperforms Diffusion Policy on real robotic tasks. GFC is shown to have lower jerk, better trajectory geometry preservation, and higher success rates, showing its potential in trajectory-based robot learning.

Critical weaknesses

The writing and organization feel unpolished and fragmented. The paper uses a large amount of mathematical notation and terms without sufficient explanation. At the sentence level, there are numerous ambiguous references and terms that make the writing hard to follow. For example, Sec. 3.4 states that "regressing these velocities yields the marginal forward field when the required conditional expectations and transport regularity hold," without specifying which conditions are required. The paper also contains inconsistencies. For example, Figure 2 shows the forward marginals approaching \mathcal{N}(0, I) without explanation, but Sec. 3.4 states that the terminal distribution is not automatically Gaussian, and the experiments fit a separate terminal source instead. Section 4.2 says it compares against Streaming Flow Policy, while Table 4 lists "Streaming Diffusion Policy".

It is unclear whether the model learns a collapsed distribution, as the training objective of the reverse model is not specified. Additionally, Figure 2 suggests that the forward process approaches \mathcal{N}(0, I), but the paper never gives a concrete derivation. The experiments use only 5 corruption steps, each driven by a small number of random coefficients, so whether this corruption reaches \mathcal{N}(0, I) remains unclear. If it does not, the reverse process may effectively start from only slightly corrupted trajectories, so most of the generative modeling would be done by fitting the terminal source, leaving a much easier task for the reverse process.

The baselines appear insufficiently tuned. On the real-robot T-block pushing and box uprighting tasks, Diffusion Policy succeeds in only 2/100 and 0/100 trials. The original Diffusion Policy paper reports a 95% success rate on a similar real-world Push-T experiment, and Table 4 of this paper reports an average score of 90.7 for Diffusion Policy on simulated Push-T, which makes the real-robot Diffusion Policy baseline questionable.

Questions for authors

1- What is the exact training objective of the reverse policy in GFC? Does this training objective ensure the learned distribution is close to the training data distribution?

2- What are the settings you used for the Diffusion Policy model on the T-block pushing and box uprighting tasks? Did you tune it on both tasks?

3- Is the terminal distribution of the forward process guaranteed to be a standard Gaussian? If so, can you justify it? If not, which source distribution is used for sampling in the robot experiments?

4- In Eq. 10, the warped metric only affects the motion through the factor e^{-\varepsilon_t S_\tau^\omega}. When \varepsilon_t is small, this factor is close to the identity, so the trajectory is mainly moved by a smooth random push under the fixed base metric. Did you try removing this factor and using only the base metric? Does GFC still perform better without it?

Comments on the paper

1 comment

Thanks for your response. So far, I don't see any revisions in the paper. As such, I will consider this concern still open and will wait for the revision to determine if the concern is resolved.

PS: please highlight the updated lines in a different color so that its clear whether something has changed.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.