EASI: Evolutionary Adversarial Simulator Identification for Sim-to-Real Transfer
Authors: Haoyu Dong, Huiqiao Fu, Wentao Xu, Zhehao Zhou, Chunlin Chen
NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In the experiments, we test EASI on 4 sim-to-sim tasks (Go2, Cartpole, Ant, Ballbalance) and 2 sim-to-real tasks (Go2, Cartpole). |
| Researcher Affiliation | Academia | Haoyu Dong, Huiqiao Fu, Wentao Xu, Zhehao Zhou, Chunlin Chen Department of Control Science and Intelligent Engineering, School of Management and Engineering, Nanjing University, China {haoyudong,hqfu, wtxu, zhzhou}@smail.nju.edu.cn, clchen@nju.edu.cn |
| Pseudocode | Yes | The schematic overview of the EASI architecture is shown in Fig. 1, and the pseudo-code is shown in Algorithm 1. |
| Open Source Code | No | Video and code are shown at our page. (This implies future availability or a project page, not direct open-source code for the described methodology readily accessible in the paper). |
| Open Datasets | No | For the collection of the demonstration dataset, we use UDR to train a rough policy, and then use the policy to control agents in the target domain collecting trajectories. (This indicates data is generated by authors and not from a publicly available source with a link/citation). |
| Dataset Splits | No | The paper mentions training policies and testing them, but it does not provide specific details on how the collected or used data is split into training, validation, and test sets with percentages or absolute sample counts. |
| Hardware Specification | Yes | In our experiment, running on a PC equipped with Intel i5-13600KF and RTX 4060 Ti, EASI completed the evolutionary adversarial searching process in less than 10 minutes. |
| Software Dependencies | No | The paper mentions using specific software components like Isaac Gym, SAC, and Ess-Info GAIL, but it does not provide specific version numbers for these software dependencies or the libraries used for neural networks. |
| Experiment Setup | Yes | We utilize Isaac Gym [39] as the simulator. In the simulator identification process, we create 300 parallel environments for EASI... we employ (µ/µI, λ)ES [44] as the generator with the setting µ = 150 and λ = 300... we set k = 3 for the Cartpole, Ant, and Ballbalance tasks, and k = 2 for Go2 task. |