REvolveR: Continuous Evolutionary Models for Robot-to-robot Policy Transfer
Authors: Xingyu Liu, Deepak Pathak, Kris Kitani
ICML 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments on a physics simulator show that the proposed continuous evolutionary model can effectively transfer the policy across robots and achieve superior sample efficiency on new robots. |
| Researcher Affiliation | Academia | 1The Robotics Institute, Carnegie Mellon University, Pittsburgh, PA 15213, USA. Correspondence to: Xingyu Liu <xingyul3@cs.cmu.edu>. |
| Pseudocode | Yes | Algorithm 1 Continuous Robot Evolution Policy Transfer |
| Open Source Code | Yes | Code is released at https: //github.com/xingyul/revolver. |
| Open Datasets | Yes | We showcase our REvolve R on three Mu Jo Co Gym environments (Brockman et al., 2016) with dense reward. We also experiment on Hand Manipulation Suite tasks (Rajeswaran et al., 2018) in sparse rewards setting. |
| Dataset Splits | No | The paper conducts experiments in simulated environments (Mu Jo Co Gym and Hand Manipulation Suite) where policies are trained and evaluated, but it does not describe using pre-defined dataset splits (e.g., train/validation/test percentages or counts) for a fixed dataset in the conventional sense. |
| Hardware Specification | No | The paper mentions using physics simulators (Mu Jo Co and pyBullet) and running experiments, but it does not provide any specific hardware details such as GPU models, CPU specifications, or memory configurations used for the experiments. |
| Software Dependencies | No | The paper mentions several software components and algorithms used (e.g., TD3, SAC, Mu Jo Co, pyBullet, NPG) but does not provide specific version numbers for any of them, which is necessary for reproducible software dependencies. |
| Experiment Setup | Yes | Experiment Hyperparameter Setting We illustrate the hyperparameters of the neural networks used in Gym and Hand Manipulation Suite experiments, including layer size, batch size and learning rate, in Table 3. |