RoboMoRe: LLM-based Robot Co-design via Joint Optimization of Morphology and Reward

Chengtian Ma; Jiawei Fang; Lining Yao; Qiuyu Lu; Yuxuan Sun

REVIEW 3 major objections 5 minor 1 cited by

RoboMoRe: LLM-based Robot Co-design via Joint Optimization of Morphology and Reward

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read RoboMoRe claims that robot co-design fails when rewards are fixed, and that jointly optimizing morphology and reward with an LLM in a coarse-to-fine loop produces dramatically more efficient robots across eight tasks.

desk verdict A promising integration of LLM reward shaping and morphology co-design, but the headline efficiency comparisons are not apples-to-apples and the results tables contradict each other. read the letter →

arxiv 2506.00276 v1 pith:FO4XHFYK submitted 2025-05-30 cs.RO cs.CL

Jiawei Fang , Yuxuan Sun , Chengtian Ma , Qiuyu Lu , Lining Yao This is my paper

classification cs.ROcs.CL

keywords robotco-designmorphologyoptimizationrewardshapinglargelanguagemodelscoarse-to-finediversityreflectionMuJoCoefficiencymetric

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that the standard co-design recipe, optimizing a robot's body while keeping a fixed reward, systematically misses the best designs because a body's best motion mode depends on its shape. RoboMoRe instead lets a large language model propose and refine morphology and reward functions together, first casting a wide net of diverse candidates and then alternating between body and reward improvements. On eight MuJoCo locomotion and jumping tasks, the framework reports efficiency (distance or height per unit volume) far above human-engineered designs and three baselines, with no task-specific prompting or reward/morphology templates. The payoff, if the results hold, is a general recipe for letting robots discover their own shapes and their own ways of moving.

What carries the argument

The load-bearing mechanism is the coarse-to-fine LLM optimization loop. In the coarse stage, a Diversity Reflection prompt tells the LLM to propose a new design that is both high-fitness and 'quite different' from all previous ones, producing diverse morphology–reward pairs that are then evaluated by reinforcement learning; this replaces the random sampling that evolutionary and Bayesian methods rely on. In the fine stage, the LLM plays the role of a gradient oracle: given top candidates and their scores, it proposes a morphology update, then a reward update, accepting each only if retraining improves efficiency, in an alternating optimization cycle. The objective that shapes all decisions is $\mathrm{efficiency} = \mathrm{fitness}/\mathrm{volume}$, computed for each robot type, which is meant to force material-aware, compact designs rather than brute-force scaling.

What would settle it

Rerun Bayesian Optimization, Eureka, and Eureka (Mor.) with $\mathrm{efficiency} = \mathrm{fitness}/\mathrm{volume}$ as the explicit optimization objective, and report their efficiency after the same training budget; if the gaps shrink to a small factor or reverse, the central claim that joint morphology–reward optimization is what drives the gains would fail. A confirmatory second check: measure the volume distributions of RoboMoRe versus baseline designs to see whether the advantage is mostly smaller bodies rather than better behavior.

Watch

Extended reading notes

Core claim

The paper's central claim is that jointly optimizing the reward function with the morphology, rather than treating reward as fixed, is what unlocks high-performance co-design. Concretely, RoboMoRe runs a coarse stage that generates 25 morphologies and 5 rewards through LLM diversity reflection, trains all 125 combinations with SAC, then takes the top 5% into a fine stage where the LLM alternately proposes morphology improvements and reward improvements, keeping only changes that raise efficiency. Reported headline numbers include Ant efficiency jumping from 68.22 (human) to 31,038.41 and Half-Cheetah from 11,975.34 to 495,373.71, with the fitness column often close to or above baselines at far smaller volumes. The paper reads these results as evidence that reward shaping lets each body find its own suited gait, such as rolling, jumping, hopping, or scuttling, and that morphology design remains essential.

Load-bearing premise

The comparison assumes every method was scored under the same objective, $\mathrm{efficiency} = \mathrm{fitness}/\mathrm{volume}$; if the baselines optimized raw fitness rather than efficiency, their high-volume designs would be penalized disproportionately in Table 2, and much of the reported gap could be an artifact of that mismatch.

Editorial extensions

If this is right

If the claim holds, fixed-reward co-design pipelines are leaving a large fraction of performance on the table, and reward shaping should become a standard axis in morphology search.
Diversity reflection should be useful beyond this setting: any LLM design loop that samples candidate structures can improve coverage by explicitly prompting against previous samples.
The fine-stage alternating update should compound with the coarse stage: ablations indicate the full pipeline beats either stage alone, most clearly in high-dimensional tasks like Half-Cheetah.
Efficiency-as-objective should generalize to material-constrained design: robots that move well per unit volume are cheaper to build, suggesting direct relevance to real-world fabrication.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The quantitative gap in Table 2 is best read as an upper bound until the baselines are re-run with efficiency as their explicit objective, since the paper only states that RoboMoRe optimized efficiency.
The 'LLM momentum' refinement is a heuristic substitute for gradients; replacing it with differentiable morphology proxies or learned surrogates could be a natural test of whether the alternating scheme or the LLM proposals carry the gain.
Efficiency as fitness over volume is a single scalar that ignores energy, control cost, and task-specific constraints; multi-objective variants could change which morphologies are preferred.
Because the prompts are deliberately general, the coarse-to-fine recipe should transfer to other generative design domains such as soft robotics or mechanisms, though that transfer is untested in this paper.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

First thing you should know: this paper proposes something genuinely useful jointly optimizing reward and morphology in an LLM loop, with a diversity-reflection trick and a coarse-to-fine search, and it demonstrates the idea on eight MuJoCo tasks. That integration is new, and the qualitative examples (long-legged ant hopping, low-profile robot rolling) are compelling. But the quantitative headline the efficiency numbers in Table 2 is not currently supported. The comparison is likely not apples-to-apples: the paper states that RoboMoRe uses efficiency (fitness/volume) as its core optimization metric, but nothing in the descriptions of Bayesian Optimization, Eureka, or Eureka (Mor.) says they optimized the same objective. If those baselines were maximizing raw fitness or task reward, then their efficiency in Table 2 was computed post hoc, and the large gaps (Ant 31,038 vs 68; Half-Cheetah 495,374 vs 12,158) could be mostly an artifact of penalizing designs that never traded off volume. The raw fitness column actually shows RoboMoRe losing to Eureka on Half-Cheetah, so the whole "significantly outperforms" claim rests on the efficiency metric.

There is also an internal inconsistency: Table 4 calls the full method RoboDesign (Full) and reports Ant 10,464.17 and Half-Cheetah 129,440.98, while Table 2 reports RoboMoRe at 31,038.41 and 495,373.71 for the same configuration. No protocol difference is explained. That either is a reporting error or suggests the runs differ in some unreported way, and either way it undermines trust in the numbers. The formal problem definition (Eq. 2) maximizes fitness, but the implementation optimizes efficiency; the authors admit this in a table note, but the formal statement and the experiments are not aligned. Also, there are no error bars or statistical tests anywhere in the main results, even though RL training is stochastic they say they average over runs, but they do not report variance.

On the positive side, the method is clearly described, the appendix prompts are detailed, and the ablation on diversity reflection is informative. The paper does not ship code or data, which makes the unreported details harder to resolve.

Bottom line: the framework is worth a look, but the empirical claim as written is not solid. A serious referee should engage, but the authors need to rerun baselines with the same efficiency objective, reconcile Tables 2 and 4, report variances, and ideally release code. I would not cite the efficiency numbers as they stand.

Referee Report

3 major / 5 minor

Summary. The paper introduces RoboMoRe, an LLM-driven framework for robot co-design that jointly optimizes morphology and reward functions. The method proceeds in two stages: a coarse stage in which an LLM with a diversity-reflection mechanism proposes morphology-reward pairs, and a fine stage in which the top candidates are refined through alternating LLM-guided morphology and reward updates. Experiments are reported across eight MuJoCo environments, with efficiency (fitness divided by volume) as the headline metric. The paper claims that RoboMoRe significantly outperforms human-designed robots and prior methods such as Bayesian optimization, Eureka, and Eureka(Mor.) on this metric.

Significance. If the reported results were fully supported, the paper would make a useful empirical contribution: it demonstrates a task-agnostic prompting scheme for LLM-based co-design, provides a diversity-reflection mechanism that improves sample diversity, includes ablations isolating reward shaping and morphology design, and reports robustness checks across gear powers and terrains. The appendix is detailed, with full prompts and examples of generated reward functions. However, the central comparison is currently undermined by an apparent mismatch between the formal optimization objective and the metric used in implementation, by inconsistent numbers between Table 2 and Table 4 for what appears to be the same configuration, and by the absence of any error bars or statistical tests. These issues directly affect the paper's headline claim of significant outperformance, so the contribution cannot be assessed as presented.

major comments (3)

[Section 5, Evaluation Metrics; Eq. (2); Table 2] The formal objective in Eq. (2) maximizes fitness F, but the implementation and Table 2 use efficiency = fitness/volume, and Section 5 states that efficiency is the core metric in the optimization iterations. The manuscript never states whether Bayesian Optimization, Eureka, and Eureka(Mor.) were run optimizing the same efficiency objective or raw fitness. If the baselines optimized raw fitness, their volume-heavy designs are penalized only at evaluation time, so the large efficiency gaps in Table 2 (e.g., Ant 31,038 vs. 68 for Human; Half-Cheetah 495,374 vs. 12,158 for Eureka) would reflect an objective mismatch rather than superior co-design. The abstract's claim of "significantly outperforms" rests entirely on this efficiency comparison, and the raw fitness column shows RoboMoRe is not best on fitness (e.g., Half-Cheetah 135.71 vs. Eureka 257.55). This must be resolved either by specifying and enforcing the same optimization objective for all methods or by demoting efficiency to a secondary metric.
[Table 2 and Table 4] Table 4 reports a row labeled "RoboDesign (Full)" that appears to denote the full RoboMoRe method, but its efficiency values differ systematically from the RoboMoRe row in Table 2 for the same tasks (e.g., Ant 10,464 vs. 31,038; Half-Cheetah 129,441 vs. 495,374; Walker 1,482.60 vs. 6,665.85). The text never defines "RoboDesign" or explains why the same configuration yields different numbers in the two tables. This is a load-bearing inconsistency: unless there is an unreported difference in training budget, selection procedure, or evaluation protocol, one of the tables is incorrect, and the reported results cannot be reproduced from the manuscript.
[Tables 2-4; Appendix D] All performance comparisons are reported as point estimates with no error bars, confidence intervals, or statistical tests, despite Appendix D stating that results are averaged over 100 independent runs. The claim that RoboMoRe "significantly outperforms" human designs and competing methods requires some measure of variability or a significance test, especially because the differences in the fitness columns are often small or in the baseline's favor. The absence of variance reporting makes it impossible to judge whether the headline efficiency advantages are robust or the result of a single favorable seed.

minor comments (5)

[Table 4 title] The table is titled "Ablation study of RoboDesign," but the method introduced in the paper is RoboMoRe; the name "RoboDesign" is never defined and should be corrected.
[Appendix E.1] The text references "Fig. ??" in the discussion of the 5 x 25 efficiency matrices; this is a broken cross-reference and should point to the actual figure.
[Appendix D, Comparison Methods] The description of Bayesian optimization, Eureka, and Eureka(Mor.) gives iteration counts and population sizes but does not specify the objective function optimized by each baseline, which is the key missing information identified in the major comments.
[Appendix A.1.2] The morphology prompt contains typos such as "Fintess function" and "eend attachment"; these should be corrected for clarity.
[Section 5, Evaluation Metrics] The volume computation is described as using "custom scripts," but no details or code are provided in the appendix; since efficiency is the central metric, the volume computation should be specified precisely or released with the code.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RoboMoRe's claims rest on external MuJoCo simulation scoring, not on fitted parameters or self-citation chains.

full rationale

The paper's derivation chain is self-contained and empirically grounded. RoboMoRe proposes morphology-reward pairs via an LLM, scores each pair by training an RL policy in MuJoCo simulation, and selects/refines candidates based on those external fitness evaluations. There is no fitted parameter that is subsequently renamed as a prediction, no equation that reduces to its own input by construction, and no load-bearing self-citation or imported uniqueness theorem. The efficiency metric (fitness/volume) is a stated, externally evaluated objective; whether the baselines optimized the same objective is an experimental comparability concern, not a circularity in the derivation. The reported fitness and efficiency numbers come from independent simulation runs, so the central comparison is not forced by the method's definitions. No circular step was found.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

RoboMoRe is an empirical pipeline without a formal derivation. The central claim rests on simulation-based fitness evaluations, fixed RL training budgets, the choice of efficiency as the objective, and the assumption that LLM proposals are genuine discoveries rather than memorized benchmark designs.

free parameters (4)

NM (number of morphologies) = 25
Chosen by hand as a speed-quality trade-off in coarse optimization (Appendix D, Table 6).
NR (number of reward functions) = 5
Chosen by hand alongside NM; affects the number of pairs trained (5 * 25 = 125).
k (fine selection fraction) = 5%
Top 5% of coarse candidates are refined; chosen empirically.
SAC training steps per candidate = 5e5 (1e6 for final)
Fixed RL training budget; if some morphologies need more steps, the rank order of candidates could change.

assumptions (4)

domain assumption SAC converges to a near-optimal policy within 5e5 steps for each morphology-reward pair.
The method ranks candidates by RL fitness after a fixed budget; Appendix D states 5e5 steps per candidate.
domain assumption MuJoCo simulation is a faithful proxy for real-world locomotion performance.
All conclusions are drawn from simulated fitness and efficiency; the paper only discusses real-world transfer as future work.
domain assumption LLM proposals are not contaminated by memorized MuJoCo benchmark designs.
The authors mask XML parameters but feed standard task descriptions, so GPT-4 may still recall near-default Ant/HalfCheetah designs; the paper does not test for this.
domain assumption Efficiency (fitness/volume) is the correct objective for comparing co-design methods.
The metric is introduced in Section 5 and drives all optimization; baselines may not have used it.

how reviews work

0 comments

Cite this review

Pith. "Pith review of RoboMoRe: LLM-based Robot Co-design via Joint Optimization of Morphology and Reward." pith.science (2026). https://pith.science/paper/FO4XHFYK

@misc{pith2026250600276,
  author       = {Pith},
  title        = {Pith review of: RoboMoRe: LLM-based Robot Co-design via Joint Optimization of Morphology and Reward},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FO4XHFYK}},
  note         = {Machine review of arXiv:2506.00276}
}

read the original abstract

Robot co-design, jointly optimizing morphology and control policy, remains a longstanding challenge in the robotics community, where many promising robots have been developed. However, a key limitation lies in its tendency to converge to sub-optimal designs due to the use of fixed reward functions, which fail to explore the diverse motion modes suitable for different morphologies. Here we propose RoboMoRe, a large language model (LLM)-driven framework that integrates morphology and reward shaping for co-optimization within the robot co-design loop. RoboMoRe performs a dual-stage optimization: in the coarse optimization stage, an LLM-based diversity reflection mechanism generates both diverse and high-quality morphology-reward pairs and efficiently explores their distribution. In the fine optimization stage, top candidates are iteratively refined through alternating LLM-guided reward and morphology gradient updates. RoboMoRe can optimize both efficient robot morphologies and their suited motion behaviors through reward shaping. Results demonstrate that without any task-specific prompting or predefined reward/morphology templates, RoboMoRe significantly outperforms human-engineered designs and competing methods across eight different tasks.

Figures

Figures reproduced from arXiv: 2506.00276 by the authors.

**Figure 1.** Diverse motion behaviors and robot morphologies generated by RoboMoRe. Orange: A bio-inspired long-legged robot crawling forward in a climber posture. Blue: A uniform robot advancing by jumping. Purple: A low-center-of-gravity robot rolling forward along the ground. Although many existing approaches have achieved promising results, they typically rely on a fixed reward function, which significantly limits the potent… view at source ↗

**Figure 2.** The overall pipeline of RoboMoRe. The LLM is provided with the task description, environment code, and masked structure file as the input prompt. The environment code specifies the simulation variables to write reward functions, while the masked structure file provides masked morphology parameters of robots for the LLM to complete. Our algorithm follows a coarse-to-fine strategy. First, the LLM leverages diversity r… view at source ↗

**Figure 3.** Overview of Coarse-to-Fine algorithm. (a) In the coarse optimization stage, the LLM first generates diverse reward functions and robot morphologies using a Diversity Reflection mechanism. These reward-morphology pairs are then trained via reinforcement learning to explore and identify the most suitable motion behavior for each robot. (b) In the fine optimization stage, the LLM employs Alternating Optimization to joi… view at source ↗

Figures from the paper (10 more)

**Figure 4.** Figure 4: Optimal designs generated by RoboMoRe. (a) Ant. (b) Ant-Powered, where motor power is doubled. (c) Ant-Desert, with the terrain replaced by desert. (d) Ant-Jump, where the task is changed from locomotion to jumping. (e) Hopper. (f) Half-Cheetah. (g) Swimmer. (h) Walker…

**Figure 5.** Figure 5: Qualitative results using three different methods: (a) Random Sample, (b) RoboMoRe (w/o Diversity Reflection), (c) RoboMoRe. It can be observed that RoboMoRe with Diversity Reflection could generate more high-quality and diverse robot morphology examples compared to ot…

**Figure 6.** Figure 6: Qualitative results of diverse motion behaviors using Diversity Reflection. (a) The short-legged, low-center-of-gravity robot rolls forward along the ground. (b) The uniform robot propels itself forward with sideways hops. (c) The crab-inspired robot repositions itself…

**Figure 7.** Figure 7: Comparison of the highest efficiency achieved in different cases among RoboMoRe, RoboMoRe (w/o fine stage), and RoboMoRe (w/o coarse stage). Both coarse and fine stages contribute significantly to co-design performance [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

**Figure 8.** Figure 8: (a-d) Motion behaviors of our four cases. (e) Time-series fitness (walk distance) across 100 independent evaluations per design case [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

**Figure 9.** Figure 9: Human designed agents in eight different environments for visualization. Purple indicates agents in an augmented gear power. C.1 Ant C.1.1 Task Description The ant is a 3D quadruped robot consisting of a torso (free rotational body) with four legs attached to it, where…

**Figure 10.** Figure 10: Coarse optimization results across 8 tasks (a–h) for Random Sample, RoboMoRe w/o Diversity Reflection (DR), and RoboMoRe. E.2 Robustness across Different Gear Powers To validate the robustness of RoboMoRe-generated designs across varying levels of motor power, we sele…

**Figure 11.** Figure 11: Different power settings and corresponding results. (a) Ant-Powered-50. (b) AntPowered-100. (c) Ant-Powered-150. (d) Ant-Powered-200. (e) Ant-Powered-250. (f) Ant-Powered300. E.3 Robustness across Different Terrains To further validate the robustness of RoboMoRe-gen…

**Figure 12.** Figure 12: Different terrain environments and corresponding results. (a) Ant-Desert. (b) AntGround. (c) Ant-Hills. (d) Ant-Snow. (e) Results for Ant-Desert. (f) Results for Ant-Ground. (g) Results for Ant-Hills. (h) Results for Ant-Snow. 24 [PITH_FULL_IMAGE:figures/full_fig_p0…

**Figure 13.** Figure 13: presents the optimized morphology design. Evidently, RoboMoRe is capable of producing highly efficient structures, demonstrating its effectiveness in morphology optimization [PITH_FULL_IMAGE:figures/full_fig_p026_13.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Debate2Create: Robot Co-design via Multi-Agent LLM Debate
cs.RO 2025-10 reject novelty 6.0 of 10

A structured multi-agent LLM debate grounded in simulation is proposed for co-designing robot morphology and reward, but the body reports only a single Ant experiment.

Reference graph

Works this paper leans on

41 extracted references · 28 canonical work pages · cited by 1 Pith paper

[1]

Spinning up in deep reinforcement learning

Joshua Achiam. Spinning up in deep reinforcement learning. 2018

work page 2018
[2]

An overview of evolutionary algorithms for parameter optimization

Thomas Bäck and Hans-Paul Schwefel. An overview of evolutionary algorithms for parameter optimization. Evolutionary computation, 1(1):1–23, 1993

work page 1993
[3]

On the use of the coefficient of variation as a measure of diversity

Arthur G Bedeian and Kevin W Mossholder. On the use of the coefficient of variation as a measure of diversity. Organizational Research Methods, 3(3):285–297, 2000

work page 2000
[4]

Convergence of alternating optimization

James C Bezdek and Richard J Hathaway. Convergence of alternating optimization. Neural, Parallel & Scientific Computations, 11(4):351–368, 2003

work page 2003
[5]

Evolution gym: A large-scale benchmark for evolving soft robots

Jagdeep Bhatia, Holly Jackson, Yunsheng Tian, Jie Xu, and Wojciech Matusik. Evolution gym: A large-scale benchmark for evolving soft robots. Advances in Neural Information Processing Systems, 34:2201–2214, 2021

work page 2021
[6]

One robot for many tasks: Versatile co-design through stochastic programming

Gabriel Bravo-Palacios, Andrea Del Prete, and Patrick M Wensing. One robot for many tasks: Versatile co-design through stochastic programming. IEEE Robotics and Automation Letters, 5(2):1680–1687, 2020

work page 2020
[7]

Learning navigation behaviors end-to-end with autorl

Hao-Tien Lewis Chiang, Aleksandra Faust, Marek Fiser, and Anthony Francis. Learning navigation behaviors end-to-end with autorl. IEEE Robotics and Automation Letters, 4(2):2007– 2014, 2019

work page 2007
[8]

GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill

Jieming Cui, Tengyu Liu, Ziyu Meng, Jiale Yu, Ran Song, Wei Zhang, Yixin Zhu, and Siyuan Huang. Grove: A generalized reward for learning open-vocabulary physical skill. arXiv preprint arXiv:2504.04191, 2025

work page Pith review arXiv 2025

Show all 41 references

[9]

Evolving rewards to automate reinforce- ment learning

Aleksandra Faust, Anthony Francis, and Dar Mehta. Evolving rewards to automate reinforce- ment learning. arXiv preprint arXiv:1905.07628, 2019

1905 arXiv
[10]

A tutorial on bayesian optimization

Peter I Frazier. A tutorial on bayesian optimization. arXiv preprint arXiv:1807.02811, 2018

2018 arXiv
[11]

Forward and inverse problems in the mechanics of soft filaments

M Gazzola, LH Dudte, AG McCormick, and L Mahadevan. Forward and inverse problems in the mechanics of soft filaments. Royal Society open science, 5(6):171628, 2018

2018
[12]

Somogym: A toolkit for developing and evaluating controllers and reinforcement learning algorithms for soft robots

Moritz A Graule, Thomas P McCarthy, Clark B Teeple, Justin Werfel, and Robert J Wood. Somogym: A toolkit for developing and evaluating controllers and reinforcement learning algorithms for soft robots. IEEE Robotics and Automation Letters, 7(2):4071–4078, 2022

2022
[13]

Reinforcement learning for improving agent design

David Ha. Reinforcement learning for improving agent design. Artificial life, 25(4):352–365, 2019

2019
[14]

Soft actor-critic algorithms and applications

Tuomas Haarnoja, Aurick Zhou, Kristian Hartikainen, George Tucker, Sehoon Ha, Jie Tan, Vikash Kumar, Henry Zhu, Abhishek Gupta, Pieter Abbeel, et al. Soft actor-critic algorithms and applications. arXiv preprint arXiv:1812.05905, 2018

2018 arXiv
[15]

Generating and evolving reward functions for highway driving with large language models

Xu Han, Qiannan Yang, Xianda Chen, Xiaowen Chu, and Meixin Zhu. Generating and evolving reward functions for highway driving with large language models. In 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), pages 831–836. IEEE, 2024

2024
[16]

Generative adversarial imitation learning

Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. Advances in neural information processing systems, 29, 2016

2016
[17]

Reward (mis) design for autonomous driving

W Bradley Knox, Alessandro Allievi, Holger Banzhaf, Felix Schmitt, and Peter Stone. Reward (mis) design for autonomous driving. Artificial Intelligence, 316:103829, 2023

2023
[18]

Evolution through large models

Joel Lehman, Jonathan Gordon, Shawn Jain, Kamal Ndousse, Cathy Yeh, and Kenneth O Stanley. Evolution through large models. In Handbook of evolutionary machine learning, pages 331–366. Springer, 2023

2023
[19]

Bodygen: Advancing towards efficient embodiment co-design

Haofei Lu, Zhe Wu, Junliang Xing, Jianshu Li, Ruoyu Li, Zhe Li, and Yuanchun Shi. Bodygen: Advancing towards efficient embodiment co-design. arXiv preprint arXiv:2503.00533, 2025. 11

2025 arXiv
[20]

Diffaqua: A differentiable computational design pipeline for soft underwater swimmers with shape interpolation

Pingchuan Ma, Tao Du, John Z Zhang, Kui Wu, Andrew Spielberg, Robert K Katzschmann, and Wojciech Matusik. Diffaqua: A differentiable computational design pipeline for soft underwater swimmers with shape interpolation. ACM Transactions on Graphics (TOG), 40(4):1–14, 2021

2021
[21]

Eureka: Human-level reward design via coding large language models

Yecheng Jason Ma, William Liang, Guanzhi Wang, De-An Huang, Osbert Bastani, Dinesh Jayaraman, Yuke Zhu, Linxi Fan, and Anima Anandkumar. Eureka: Human-level reward design via coding large language models. arXiv preprint arXiv:2310.12931, 2023

arXiv
[22]

Learning to control self-assembling morphologies: a study of generalization via modularity

Deepak Pathak, Christopher Lu, Trevor Darrell, Phillip Isola, and Alexei A Efros. Learning to control self-assembling morphologies: a study of generalization via modularity. Advances in Neural Information Processing Systems, 32, 2019

2019
[23]

Robomorph: Evolving robot morphology using large language models

Kevin Qiu, Krzysztof Ciebiera, Paweł Fijałkowski, Marek Cygan, and Łukasz Kuci ´nski. Robomorph: Evolving robot morphology using large language models. arXiv preprint arXiv:2407.08626, 2024

2024
[24]

Gaussian processes in machine learning

Carl Edward Rasmussen. Gaussian processes in machine learning. In Summer school on machine learning, pages 63–71. Springer, 2003

2003
[25]

Intelligent robotic control

George Saridis. Intelligent robotic control. IEEE Transactions on Automatic Control, 28(5):547– 557, 1983

1983
[26]

Evolving virtual creatures

Karl Sims. Evolving virtual creatures. In Seminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 699–706. 2023

2023
[27]

Jasper Snoek, Hugo Larochelle, and Ryan P. Adams. Practical bayesian optimization of machine learning algorithms, 2012

2012
[28]

Laser: Towards diversified and generalizable robot design with large language models

Junru Song, Yang Yang, Huan Xiao, Wei Peng, Wen Yao, and Feifei Wang. Laser: Towards diversified and generalizable robot design with large language models. In The Thirteenth International Conference on Learning Representations
[29]

The artificial life route to artificial intelligence: Building embodied, situated agents

Luc Steels and Rodney Brooks. The artificial life route to artificial intelligence: Building embodied, situated agents. Routledge, 2018

2018
[30]

Gymnasium: A standard interface for reinforcement learning environments

Mark Towers, Ariel Kwiatkowski, Jordan Terry, John U Balis, Gianluca De Cola, Tristan Deleu, Manuel Goulão, Andreas Kallinteris, Markus Krimmel, Arjun KG, et al. Gymnasium: A standard interface for reinforcement learning environments. arXiv preprint arXiv:2407.17032, 2024

2024 arXiv
[31]

Neural graph evolution: Towards efficient automatic robot design

Tingwu Wang, Yuhao Zhou, Sanja Fidler, and Jimmy Ba. Neural graph evolution: Towards efficient automatic robot design. arXiv preprint arXiv:1906.05370, 2019

1906 arXiv
[32]

Softzoo: A soft robot co-design benchmark for locomotion in diverse environments

Tsun-Hsuan Wang, Pingchuan Ma, Andrew Everett Spielberg, Zhou Xian, Hao Zhang, Joshua B Tenenbaum, Daniela Rus, and Chuang Gan. Softzoo: A soft robot co-design benchmark for locomotion in diverse environments. arXiv preprint arXiv:2303.09555, 2023

2023 arXiv
[33]

Towards socially and morally aware rl agent: Reward design with llm

Zhaoyue Wang. Towards socially and morally aware rl agent: Reward design with llm. arXiv preprint arXiv:2401.12459, 2024

2024 arXiv
[34]

Multi-objective graph heuristic search for terrestrial robot design

Jie Xu, Andrew Spielberg, Allan Zhao, Daniela Rus, and Wojciech Matusik. Multi-objective graph heuristic search for terrestrial robot design. In 2021 IEEE international conference on robotics and automation (ICRA), pages 9863–9869. IEEE, 2021

2021
[35]

Text2reward: Reward shaping with language models for reinforcement learning

Yanchao Yang. Text2reward: Reward shaping with language models for reinforcement learning. In International Conference on Learning Representations (ICLR), 2024 (07/05/2024-11/05/2024, Vienna, Austria), 2024

2024
[36]

Cuda-accelerated soft robot neural evolution with large language model supervi- sion

Lechen Zhang. Cuda-accelerated soft robot neural evolution with large language model supervi- sion. arXiv preprint arXiv:2405.00698, 2024

2024 arXiv
[37]

Modeling and simulation of complex dynamic musculoskeletal architectures

X Zhang, FK Chan, T Parthasarathy, and M Gazzola. Modeling and simulation of complex dynamic musculoskeletal architectures. Nature Communications, 10(1):1–12, 2019. 12

2019
[38]

Robogrammar: graph grammar for terrain-optimized robot design

Allan Zhao, Jie Xu, Mina Konakovi´c-Lukovi´c, Josephine Hughes, Andrew Spielberg, Daniela Rus, and Wojciech Matusik. Robogrammar: graph grammar for terrain-optimized robot design. ACM Transactions on Graphics (TOG), 39(6):1–16, 2020

2020
[39]

def _get_rew():

Yaoming Zhu, Sidi Lu, Lei Zheng, Jiaxian Guo, Weinan Zhang, Jun Wang, and Yong Yu. Texygen: A benchmarking platform for text generation models. In The 41st international ACM SIGIR conference on research & development in information retrieval, pages 1097–1100, 2018. 13 Appendix...

2018
[40]

For reducing material cost to ensure the effieciency of robot design, you should reduce redundant paramters and increase paramters who control the robot
[41]

parameters

Your design should fit the control gear and others parts of robots well. Output Format: Please output in json format without any notes: { "parameters": [<param1>, <param2>, ..., <param10>], "description": "<your simple design style description>", } # param1 is the size of the ...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Spinning up in deep reinforcement learning

Joshua Achiam. Spinning up in deep reinforcement learning. 2018

work page 2018

[2] [2]

An overview of evolutionary algorithms for parameter optimization

Thomas Bäck and Hans-Paul Schwefel. An overview of evolutionary algorithms for parameter optimization. Evolutionary computation, 1(1):1–23, 1993

work page 1993

[3] [3]

On the use of the coefficient of variation as a measure of diversity

Arthur G Bedeian and Kevin W Mossholder. On the use of the coefficient of variation as a measure of diversity. Organizational Research Methods, 3(3):285–297, 2000

work page 2000

[4] [4]

Convergence of alternating optimization

James C Bezdek and Richard J Hathaway. Convergence of alternating optimization. Neural, Parallel & Scientific Computations, 11(4):351–368, 2003

work page 2003

[5] [5]

Evolution gym: A large-scale benchmark for evolving soft robots

Jagdeep Bhatia, Holly Jackson, Yunsheng Tian, Jie Xu, and Wojciech Matusik. Evolution gym: A large-scale benchmark for evolving soft robots. Advances in Neural Information Processing Systems, 34:2201–2214, 2021

work page 2021

[6] [6]

One robot for many tasks: Versatile co-design through stochastic programming

Gabriel Bravo-Palacios, Andrea Del Prete, and Patrick M Wensing. One robot for many tasks: Versatile co-design through stochastic programming. IEEE Robotics and Automation Letters, 5(2):1680–1687, 2020

work page 2020

[7] [7]

Learning navigation behaviors end-to-end with autorl

Hao-Tien Lewis Chiang, Aleksandra Faust, Marek Fiser, and Anthony Francis. Learning navigation behaviors end-to-end with autorl. IEEE Robotics and Automation Letters, 4(2):2007– 2014, 2019

work page 2007

[8] [8]

GROVE: A Generalized Reward for Learning Open-Vocabulary Physical Skill

Jieming Cui, Tengyu Liu, Ziyu Meng, Jiale Yu, Ran Song, Wei Zhang, Yixin Zhu, and Siyuan Huang. Grove: A generalized reward for learning open-vocabulary physical skill. arXiv preprint arXiv:2504.04191, 2025

work page Pith review arXiv 2025