Transform2Act: Learning a Transform-and-Control Policy for Efficient Agent Design

Authors: Ye Yuan, Yuda Song, Zhengyi Luo, Wen Sun, Kris M. Kitani

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments show that our approach, Transform2Act, outperforms prior methods significantly in terms of convergence speed and final performance.
Researcher Affiliation Academia 1Carnegie Mellon University, 2Cornell University {yyuan2,yudas,zluo2,kkitani}@cs.cmu.edu, ws455@cornell.edu
Pseudocode Yes An overview of our method is provided in Figure 2. We also outline our approach in Algorithm 1.
Open Source Code Yes Code and videos are available at https://sites.google.com/view/transform2act.
Open Datasets No The paper uses simulation environments (Mu Jo Co simulator) for experiments rather than external, publicly available datasets. Therefore, there is no dataset for which to provide concrete access information.
Dataset Splits No The paper uses a simulation environment, and while it discusses training, it does not specify traditional dataset splits (e.g., percentages or counts for training, validation, and test sets) that would be applicable to a fixed dataset. Instead, it refers to batch sizes and training epochs within the reinforcement learning paradigm.
Hardware Specification Yes For all the environments used in the paper, it takes around one day to train our model on a standard server with 20 CPU cores and an NVIDIA RTX 2080 Ti GPU.
Software Dependencies No The paper mentions using
Experiment Setup Yes In this section, we present the hyperparameters searched and used for our method in Table 1 and the hyperparameters for the baselines in Table 2.