acceptodds
Under review as a conference paper at ICLR 2027

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

Abstract

Robot co-design, the joint optimization of morphology and control policy, remains a longstanding challenge in the robotics community. Existing approaches often converge to suboptimal designs because they rely on fixed reward functions, which fail to capture the diverse motion modes suited to different morphologies. We propose RoboMoRe, a large language model (LLM)-driven framework that integrates morphology and reward shaping for co-optimization within the robot design loop. RoboMoRe adopts a dual-stage strategy: in the coarse stage, an LLM-based Diversity Reflection mechanism is proposed to generate diverse and high-quality morphology–reward pairs and Morphology Screening is performed to reduce unpotential candidates and efficiently explore the design space; in the fine stage, top candidates are iteratively refined through alternating LLM-guided updates to both reward and morphology. This process enables RoboMoRe to discover efficient morphologies and their corresponding motion behaviors through joint optimization. The result across eight representative tasks demonstrate that without any task-specific prompting or predefined reward and morphology templates, RoboMoRe significantly outperform human-engineered design results and competing methods. Additional experiments demonstrate robustness of RoboMoRe on manipulation and free-form design tasks.

Primary area:
applications to robotics, autonomy, planning
Keywords:
LLM, Robot Co-design

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_4idTOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an LLM based framework for jointly optimize robot morphology and reward functions. It designs a coarse-to-fine optimization strategy, including diversity reflection, morphology screening, and alternative optimization. Then the design is evaluated using SAC, an RL algorithm. Experiments on different robotics tasks demonstrate promising improvements.

Critical strengths
  1. Joint optimization of morphology and reward function can indeed find better solution than single optimization.
  2. This paper demonstrated the effect on different kinds of robots and provide good visualizations.
Critical weaknesses
  • The description is a little misleading. In Figure 3, there is a ∂LLMf/∂θ\partial_{LLM}f/\partial_\theta. However, this paper does not actually calculate gradients. Instead, it simply asks the LLM to propose better parameters based on previous results. This is more like prompt engineering rather than a truly differentiable joint optimizer for morphology design and reward functions.
  • Combining morphology optimization and reward function optimization seems relatively straightforward. The performance improvements are somewhat expected even without these experiments.
  • Since existing works such as LASeR (LLM for morphology design) and Eureka (LLM for reward shaping) have already explored these directions, I feel this paper mainly combines these two approaches without introducing a significant methodological breakthrough.
Questions for authors
  • What is the key methodological novelty beyond combining existing LLM-based design approaches?
  • Could the authors compare against LASeR combined with Eureka?
  • How are the proposed LLM "textual gradients" defined? Are they actual gradient estimates or simply heuristic proposals?
Reviewer_wEtSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This work proposes RoboMoRe, a robot co-design framework that jointly optimizes morphology and reward functions. In the coarse stage, the LLM generates diverse reward functions and robot morphologies conditioned on previously generated samples and guided by diversity-promoting prompts. After control policy training through RL, unpromising morphologies are filtered out. In the fine stage, the LLM alternately refines morphology and reward functions, training and evaluating a control policy after each update. A morphology–reward pair is evaluated using efficiency in the framework, defined as task fitness divided by robot volume. A key contribution is to explicitly include LLM-generated reward functions in the search space alongside robot morphology. Experiments demonstrate that the framework improves task performance per unit robot volume, validate the contributions of its components, and show its applicability across diverse simulation tasks.

Critical strengths
  1. The joint optimization of robot morphology and reward functions is an interesting and relatively novel idea.
  2. The paper provides extensive experimental evaluation, including baseline comparisons, ablation studies, and additional simulation tasks.
Critical weaknesses

While jointly searching robot morphologies and reward functions is an interesting contribution, the optimization procedure itself largely relies on established strategies, including history-conditioned prompting, candidate screening, and alternating optimization. The main concern is that the experiments do not sufficiently establish the specific benefit of LLM-based feedback-driven refinement beyond joint search and alternating optimization. To clarify the benefit of LLM-guided refinement, the authors should compare it with generating candidates without feedback and with non-LLM optimization methods, using the same alternating procedure and evaluation budget. It is also unclear how reliably the LLM can identify useful changes from performance feedback, which may reflect differences in morphology, reward design, or randomness in policy training.

Questions for authors

This work uses efficiency as the primary optimization objective to encourage material-aware design rather than brute-force scaling. Could the authors clarify the physical interpretation and practical significance of efficiency? What meaningful aspect of robot performance does this ratio capture, and why is it an appropriate primary objective for robot co-design?

Reviewer_2v2ROpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

A LLM was used to co-optimize the size of a given robot's body parts and its reward function.

Critical strengths

The paper shows bodies with different sized parts need different reward functions.

Critical weaknesses

The body can be stretched and smushed but it is always the same basic structure. The LLM yields a stick figure that is as large and light as possible. This is usually worse than the baselines in terms of performance. It obviously has higher energy density, but this was not the objective.

The role of prior knowledge is unclear. The LLM receives familiar robot names, structures, and task descriptions. Masking numerical parameters does not remove these priors. Testing unfamiliar structures, such as a three-legged Ant, would help assess whether the method can go beyond familiar robot designs.

The benefit of reward optimization on EvoGym is unclear. EvoGym provides a more flexible design space, but there is no ablation comparing the full method with a version using fixed rewards.

Screening with only one reward seems to contradict the paper’s main motivation that different morphologies may need different rewards. A morphology that performs poorly with the first reward may perform well with another, but screening may discard it before this can be tested.

Questions for authors

A simple joint optimization baseline is missing. For example, how does RoboMoRe compare with BO for morphology optimization combined with LLM-based reward search?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.