acceptodds
Under review as a conference paper at ICLR 2027

Learning Transfer-Aware Environment Curricula from Cross-Simulator Feedback

Abstract

A harder navigation task need not offer more useful training for a physical robot. CrossSim-Curriculum selects geometry edits for cross-floor quadruped navigation by predicting their physical learning gain. Paired short updates in two simulators compare each edit with a control from the same checkpoint. A residual model calibrates these gains with sparse physical feedback; selection then accounts for uncertainty, cost, and edit-family coverage. We derive adaptive calibration intervals and local selection bounds, with seed, horizon, and route-shift allowances connecting probes to deployment. The evaluation protocol tests edit ranking and policy transfer across geometry splits, a held-out backend, and physical routes at matched budgets.

Primary area:
reinforcement learning
Keywords:
curriculum learning, sim-to-real transfer, environment design, robot learning

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_X3vQOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes CROSSSIM-CURRICULUM, a method for selecting environment edits according to their predicted benefit for physical robot navigation. For each candidate edit, it performs short edited-versus-control training probes in two simulators from a shared policy checkpoint. The simulator gains, their disagreement, source difficulty, and edit descriptors are used to predict physical learning gain, with a Bayesian residual model calibrated using sparse physical trials. The curriculum selector accounts for predicted gain, uncertainty, training cost, and coverage across edit families. The paper presents conditional calibration and local-selection bounds and reports comparisons on structural and compositional navigation splits, a held-out simulator, and physical routes. However, the experimental reporting and supporting material are incomplete.

Critical strengths

The central objective is meaningful: an environment that is difficult in simulation is not necessarily useful for improving physical performance. Comparing edited and control training runs from the same checkpoint provides a sensible way to estimate the incremental value of an environment edit.

Critical weaknesses

The submission is incomplete to an extent that prevents a reliable assessment of its results. It refers to Appendices A–D for proofs, implementation details, cost accounting, and diagnostic experiments, but none are included, and the linked repository is empty. Section 6 still describes a proposed test plan rather than clearly documenting the experiments underlying Table 1. Essential information, including the robot and controller, simulator configurations, actual seed and site counts, and numerical training budgets, is missing. This goes beyond presentation quality: the reported physical-transfer results cannot be adequately checked or reproduced.

Questions for authors

Which engines correspond to simulators A and B, and what configurations differ between them? Please specify how the robot, control interface, observations, and rewards are matched, together with the probe horizon K, student-training horizon L, and total simulation budget.

Reviewer_GDt5OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

My understanding is limited because this paper is very hard to make sense of (see weaknesses). However, what I did understand was that the authors present a method to select edits to the geometry of a quadruped navigation environment (e.g., stair width, landing size, clutter) by predicting how much training on each edit will improve navigation performance. They do this by training copies of the current policy in two simulators (Isaac Lab and MuJoCo), once on the edited scene and once on the unedited scene, and measuring the difference in performance. These simulator gains, along with the edit's difficulty and a descriptor of the edit, are fed into a Bayesian linear model that predicts the gain on the physical robot (unclear how they get real physical robot targets). Unfortunately, due to the writing quality, it is unclear to me exactly what the experiments are and show, why these contributions matter, and whether the experiments and theoretical analysis adequately support the claims made.

Critical strengths

Assuming their contributions are real, the authors provide a method for estimating how much training on a given environment edit will improve a physical robot's navigation performance. This seems like an important problem for curriculum design, since, as the authors point out, a harder environment is not necessarily a more useful one to train on.

Critical weaknesses

The paper reads as AI-generated at worst, or as unpolished and incomplete at best. For starters, the header at the top of the paper reads "Manuscript for advisor discussion." The introduction is full of undefined jargon. For example:

  • What is an "equal-budget control"?
  • What is "edit-specific evidence"?
  • What is a "residual correction from sparse physical feedback"? A residual of what? And what is meant by "sparse physical feedback"?

After reading the introduction, I have no idea what the problem is, why it matters, or what their contribution is.

Other sections are just as problematic. For example, the related work section lists references but never describes how their method improves on and/or builds on the existing literature. The experimental section reads like a plan, not something that was actually done, e.g., "The test plan allows at most five policy seeds [...]" and "Pilot power and site access determine the final plan before testing." And in Table 1, the authors compare against 13 baselines, most of which are only referenced in the related work section, and only two are ever described in the text (A-only + real and real edit B0), and even then, they're still not very clear. Some, e.g., "SimOpt-style" and "Multi-sim uniform," are never referenced except on the table!

Questions for authors
  1. Can you clearly state the problem you are addressing, why it matters, and precisely how your method differs from existing work on curriculum design? Which experiments or theoretical results in the paper support these claims?
  2. Section 6 reads like an experimental plan (see weaknesses). Were any experiments actually run?
Reviewer_6YzkOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes CrossSim-Curriculum, a method for selecting geometry edits (stair width, landing size, clutter, occlusion) for training a quadruped on cross-floor navigation. Instead of using source-simulator difficulty, it estimates each edit's physical learning gain relative to an equal-budget control. Paired short training updates in two simulators (Isaac Lab/PhysX and MuJoCo) provide features, and a Bayesian ridge model fit to sparse real-robot labels calibrates the predicted gain. Edits are selected by a score combining the calibrated gain, an uncertainty penalty, cost, and a family-coverage bonus.

Critical strengths

First, the paper targets a real gap: scene difficulty in a source simulator is an imperfect proxy for what a physical robot learns. Using cross-simulator disagreement as a signal, and calibrating it with a small amount of real data, is a sensible and practical way to bring sim-to-sim information into sim-to-real transfer.

Second, interesting comparison that spans curriculum, regret-based, domain-randomization, and sim-to-real methods.

Critical weaknesses

The motivation and the gap are not clearly presented. This is my main concern. The idea may well be useful, but the paper never clearly explains why it was needed or what gap in existing work it fills. The introduction is only a few short paragraphs, with no extended version in an appendix. Related work is mostly lists of citations with little explanation of how each line of work falls short for this problem. The gap is only hinted at in a single sentence ("Our selection target is the physical gain attributable to a geometry edit"). Because of this, it is hard to follow what the core contribution is and why the baselines were chosen. In its current form, I consider this a major drawback.

The baselines are not clearly described. Several baselines are labelled "-style" (SimOpt-style, BayRn-style), which suggests the authors adapted ideas from other papers to their setting. The main text does not say how this was done: what was kept, what was changed, and how they were tuned and budgeted. Without this, the comparison cannot be judged as fair, and the results in Table 1 are much weaker as evidence.

The gains are reported on the mean, but the variance is high and no statistical test is given. Table 1 has seed SDs of 10-20 points on the structural and compositional splits, for both baselines and the proposed method (16.0 and 17.1 for the full method), with at most five seeds. The improvements over the strongest baselines and the matched-budget controls (for example, 73.3 ± 4.8 vs. 67.5 ± 11.2 for PAIRED on real, and 78.2 ± 4.4 vs. 76.3 ± 5.7 for A-only + real on Sim-C) are small compared to this spread. Because of this, judging the method only by the mean is not a fair way to present the results. The protocol promises hierarchical confidence intervals, yet none are reported in the main text, and the shaded bands in Figure 3 are seed SDs, not confidence intervals. I would ask the authors to add a statistical analysis showing that the method is significantly better than the baselines, for example paired tests or bootstrap confidence intervals on the per-seed differences, with a correction for comparing against many baselines. Without this, it is not possible to tell whether the reported gains are real or due to seed noise.

Key content is in an appendix that is not part of the submission. The main text repeatedly points to appendices for the proofs, the terminal-guarantee condition, the training procedure and ledgers, and the mechanism ablations and cost accounting (Tables 2 and 3). None of these are included, so important claims cannot be verified from what was submitted.

The code is not available. The paper says code and supporting material are in an anonymous repository. At the time of my review, the repo is still empty, so the results cannot be checked or reproduced. If the material is meant for reviewers, it should be provided during the review period. The authors could use https://anonymous.4open.science/ to share code anonymously.

Questions for authors
  1. Questions: Please revise the introduction and related work to state the motivation and the gap clearly. What specific limitation of existing work (curriculum/environment-design methods, domain randomization, and sim-to-real feedback methods) does CrossSim-Curriculum address, and how does it differ from each of these? A short extended version of both sections, in the main text or an appendix, would help readers see the core contribution and why these baselines were chosen.

  2. Because the variances in Table 1 are high for both the baselines and the proposed method, please add one or two statistical significance analyses, for example paired tests or bootstrap confidence intervals on the per-seed differences, with a correction for multiple comparisons. This would show whether the improvements over the baselines and the matched-budget controls are real or within seed noise.

  3. Please push the code and supporting material to the anonymous repository before the end of the rebuttal period, so that reviewers can examine them. The repo is currently empty.

  4. Describe how the "-style" baselines (SimOpt-style, BayRn-style) were implemented, tuned, and budgeted.

  5. Provide the appendices (proofs, ablations, cost accounting), or at least the key ablation tables, so the claims about simulator B and physical calibration can be checked.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.