acceptodds
Withdrawn from ICLR 2027

Towards Scaling Search in Robot Co-Design

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Robot co-design has begun to scale its search to millions of candidate bodies. Search methods that train a controller for every body can scale by screening candidates with brief training, keeping the highest-scoring body and retraining it fully. We measure the screen's exact gain over a random body as the screen budget, candidate count (up to 128) and total compute vary. Each body in our pools has separately trained screening and final controllers. In three of four design spaces, keeping the best of four after a quarter of the full training budget improves the final score by 0.63 to 0.73 standard deviations over a random body. More candidates add to the gain, and at the same total compute the screen beats training two bodies fully and retraining the better one by 0.27 to 0.29 standard deviations. In a Walker2d variant whose reward pays only for forward travel, the same screen picks bodies that travel about 5% less than a random body. A preregistered test confirmed this harm, and it replicated with new seeds and grew with longer final training. Brief training favors heavy walkers that are weak for their weight, which full training ranks low. The harm fades with more candidates. Continuing this screen's controllers costs less and beats retraining in every pool we trained. In the walker, keeping the best of four after half the training budget instead improves the final score by 0.09 to 0.11 standard deviations over a random body. We release all our pools and code at https://anonymous.4open.science/r/scaling-codesign-search, and any best-of-N screen at our recorded budgets can be scored on the pools without training.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot co-design, morphology optimization, reinforcement learning, best-of-N selection, multi-fidelity optimization

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_H27YOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

In this paper, the authors measure the expected final-performance gain of best-of-N morphology screening on fixed pools, using independent screening and final controller seeds. It studies seven trained pools across four locomotion design spaces and supplements them with an EvoGym archive. Short screening usually helps, but a forward-travel Walker2d variant exhibits replicated negative selection lift. The contribution is an empirical dataset and characterization of screening failure.

Critical strengths
  1. The paper provides a credible negative-result analysis, which includes independent final seeds, new bodies, a second seed set, longer training, and a held-out termination-rule test make the Walker reversal substantially more convincing than a selected learning-curve anecdote. The authors correctly treat overlapping seed splits as dependent and report simultaneous bands.

  2. Useful empirical extension of an existing problem. Arza et al. already study co-design training-budget allocation, and Mertan and Cheney show undervaluation of morphological potential. This paper's additional value is the measured selection lift, independent retraining protocol, and systematic budget/candidate sweeps, not discovery that early and late fitness can disagree.

Critical weaknesses
  1. I think the released per-body learning curves and negative results could materially improve evaluation of low-fidelity morphology selection. The contribution is narrower than general brain-body co-design scaling. Moreover, generalization to other learners and deployable morphology quality remains unestablished.

  2. The observed failure remains specific to a weak, fixed learning/evaluation regime (§3.2, Tables 2, 10–11).** All new pools use one PPO configuration with unnormalized observations; Walker evaluation is a single deterministic episode from an unperturbed start, final scores are still improving, and 98% of final episodes end in a fall. Repeating seeds and termination rules confirms this pipeline's failure, but does not show that the screened bodies have intrinsically inferior achievable control performance. The mass/torque analysis is observational, with action and exploration scales tied to actuator gear. The supported conclusion is “short training misranks designs under this learner and objective,” not a causal account of body potential or a general half-budget safety rule. This limits significance rather than invalidating the finite-budget result.

  3. Deployment-facing compute claims are less complete than the selection analysis . The full-budget comparator retrains the winner even after screening at H; continuation reuses the selected training runs and therefore changes both cost and the target from reproducible body quality to a selected body-controller pair. The paper acknowledges this distinction, but the practical optimal allocation is not established by the retraining-only comparison or the post hoc envelope. Environment steps also omit morphology-dependent runtime. The retrained pools reach N=128; the million-body archive uses a different controller optimizer and continuation, so it cannot establish large-scale retraining or iterative-search behavior. Tighten the practical recommendation to these precisely defined protocols.

Questions for authors
  1. Is the intended target reproducible performance after fresh training, or best retained body-controller performance? Explain why the full-budget comparison retrains rather than retains already-trained controllers.

  2. Do any existing records include perturbed starts, observation normalization, or another PPO configuration? If not, explicitly delimit the Walker conclusion to the stated learner and fixed-start objective; no new large experiment is requested for rebuttal.

  3. Supply accessible timestamped preregistrations/amendments and the bootstrap replicate settings. Which practical budget recommendations were fixed before results, rather than selected from the same grid?

Reviewer_Dd5nOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
3
Read review
Paper summary

This paper systematically investigates the effectiveness of low-cost morphology screening in robot co-design across different candidate counts and training budgets. The results show that screening generally improves search efficiency, while identifying a Walker case where early training performance leads to poor morphology selection. The authors provide extensive experiments and detailed analyses to explain these findings.

Critical strengths
  1. The experiments are thorough and carefully designed, covering multiple design spaces with preregistered tests, replications on new bodies and seeds, and rigorous statistical reporting.
  2. The released pools, with per-seed returns at twelve budgets, together with the code, provide a useful resource for evaluating screening strategies without retraining.
Critical weaknesses
  1. Insight beyond prior work. The paper follows prior settings and confirms that screening is effective, largely consistent with existing results. I suggest the authors consider analyzing when screening becomes reliable, e.g., whether the controller has entered an effective behavioral phase (such as completing a gait cycle), using signals like episode length, fall rate, or rank stability across checkpoints, rather than a fixed fraction of the final budget.
  2. Limited generality. All tasks reward forward travel on flat ground and share a single PPO configuration. The only significant negative case comes from the walker, where 98% of final episodes still end in a fall at the full budget. This limits how far the conclusions explain screening behavior in general co-design problems.
Questions for authors

The title suggests a study of scaling in co-design search, but the paper only examines single-step best-of-N selection on fixed pools, does not study iterative search or scaling behavior. Could the authors clarify in what sense the work addresses scaling, or consider a title closer to the actual scope?

Reviewer_YBzJOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper explores a major open challenge in robot morphology and controller co-design: that a good body may be perceived as poor unless paired with a good control policy. As such, it remains an open question of how to allocate optimization budgets for robot body exploration and control policy optimization with a pre-existing design.

This paper proposes to evaluate the capacity of a candidate robot's body through reward it achieves after a short training budget. This initial evaluation period is used to select the best design which is granted a full control policy optimization period.

This paper has serious weaknesses related to their control experiments having substantially lower environment interaction budget than their proposed method.

Critical strengths

None

Critical weaknesses

Throughout the paper, almost every experiment lacks proper experimental controls at the most fundamental level. The proposed method is compared against a baseline that is not given a fair environmental interaction budget. In the worst case the proposed method is given ~20000x more environmental interactions (Fig. 3f). This paper does seem cognizant that this is an issue and it does report a single experiment with a balanced interaction budget (called 'equal compute' in Sect. 4.3; Fig. 4(e); Sect. 6). However, throughout every metric (the 'lift') is relative to an unbalanced optimization budget.

Design space, baselines, and scaling. At best, this paper introduces a 'screening' based random search method, and compares against a random search baseline. While this is an entirely valid search method, it is widely recognized as intractable for high-dimensional search spaces. Other methods tackle the problem of identifying good areas of design space in an entirely different manner including novelty search, and incremental refinement methods (e.g. CMA-ES) The paper should discuss these methods either with them as baselines, or to justify why they are not relevant for comparison.

This paper acknowledges that the early-training screen is ineffective for designs which are hard to control (e.g. walker results falling over). These types of designs are especially common in high-dimensional search spaces. While the paper does utilize the robo-grammar design space, it only evaluates a few candidate designs in that space which is a vanishingly small fraction. For the EvoGym archive analysis (which explores 1m designs), the results mirror that of the walker -- low-training budget screens are worse than selecting a random design and training to completion (even when using 17x total eval budget!)

Questions for authors

None

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

Related papers

Discussion (0)

Sign in to comment.