LASeR: Towards Diversified and Generalizable Robot Design with Large Language Models
Authors: JUNRU SONG, Yang Yang, Huan Xiao, Wei Peng, Wen Yao, Feifei Wang
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our simulated experiments on voxel-based soft robots showcase distinct advantages of LASe R over competitive baselines. Code at https://github.com/Woody SJR/LASe R. We begin this section with an introduction to our experimental setups, and then analyze the results of our comparison and ablation studies in detail. |
| Researcher Affiliation | Academia | 1School of Computer Science, Shanghai Jiao Tong University 2Intelligent Game and Decision Laboratory 3School of Statistics, Renmin University of China 4Center for Applied Statistics, Renmin University of China EMAIL, EMAIL EMAIL EMAIL |
| Pseudocode | Yes | P PSEUDO CODE OF LASER Algorithm 1: LASe R: LLM-Aided Evolutionary Search for Robot Design Automation Input: A task instance T, maximum number of evaluations M, population size N, survival rate r, maximum number of LLM interactions in each generation L, probability of similarity check p, similarity threshold s. Output: M robot designs together with their fitness scores. |
| Open Source Code | Yes | Our simulated experiments on voxel-based soft robots showcase distinct advantages of LASe R over competitive baselines. Code at https://github.com/Woody SJR/LASe R. Our code is readily available on Git Hub. |
| Open Datasets | Yes | We base our experiments on Evolution Gym (Evo Gym; Bhatia et al., 2021), a simulation environment designed for voxel-based soft robots (VSRs). For more information on Evo Gym, please refer to Bhatia et al. (2021). |
| Dataset Splits | No | The paper uses Evolution Gym, a simulation environment, for experiments, where robot designs are generated and evaluated rather than using a pre-existing dataset with fixed training/test/validation splits. There is no mention of dataset splitting for the experimental setup. |
| Hardware Specification | Yes | Our experiments are conducted on a server equipped with Intel Xeon processors running at 2.20 GHz and four NVIDIA Tesla RTX GPUs, with the system operating under Ubuntu 22.04. |
| Software Dependencies | Yes | We use GPT-4o-mini for both LASe R and LLM-Tuner, with the temperature parameter set as 0.7. For ablation studies, we additionally try out GPT-3.5-Turbo and temperatures of 1 and 1.5. Our experiments are conducted on a server equipped with Intel Xeon processors running at 2.20 GHz and four NVIDIA Tesla RTX GPUs, with the system operating under Ubuntu 22.04. This work uses GPT-3-turbo and GPT-4o-mini, whose APIs are publicly accessible. |
| Experiment Setup | Yes | Implementation Details. We use GPT-4o-mini for both LASe R and LLM-Tuner, with the temperature parameter set as 0.7. For ablation studies, we additionally try out GPT-3.5-Turbo and temperatures of 1 and 1.5. Following the common practice in previous VSR studies (Song et al., 2024a; Saito & Oka, 2024; Dong et al., 2023; Bhatia et al., 2021), we choose the simple yet effective control protocol for fitness evaluation, i.e., Multilayer Perceptron (MLP) as the controller for each robot design and PPO algorithm (Schulman et al., 2017) for policy training. Following previous studies on VSR design (Song et al., 2024a; Saito & Oka, 2024; Dong et al., 2023; Bhatia et al., 2021), robot designs are constrained to a 5 x 5 bounding box for an expressive yet tractable search space. For fair comparison, each method is permitted 1000 robot evaluations. Appendix C: HYPERPARAMETER SETTINGS. |