Combining Model-Based and Model-Free Updates for Trajectory-Centric Reinforcement Learning

Authors: Yevgen Chebotar, Karol Hausman, Marvin Zhang, Gaurav Sukhatme, Stefan Schaal, Sergey Levine

ICML 2017 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Our simulation and real-world experiments demonstrate that this method can solve challenging manipulation tasks with comparable or better performance than model-free methods while maintaining the sample efficiency of model-based methods.
Researcher Affiliation Academia 1University of Southern California, Los Angeles, CA, USA 2Max Planck Institute for Intelligent Systems, T ubingen, Germany 3University of California Berkeley, Berkeley, CA, USA.
Pseudocode Yes Algorithm 1 PILQR algorithm
Open Source Code Yes The performance of each method can be seen in our supplementary video.2 https://sites.google.com/site/icml17pilqr
Open Datasets Yes The reacher task from Open AI gym (Brockman et al., 2016)
Dataset Splits No The paper mentions 'test conditions' for evaluation but does not specify the training, validation, or testing split percentages or counts for the datasets used.
Hardware Specification Yes To evaluate our method on a real robotic platform, we use a PR2 robot (see Figure 1) to learn the following tasks:
Software Dependencies No The paper does not provide specific version numbers for any software dependencies used in the experiments.
Experiment Setup Yes Additional experimental setup details, including the exact cost functions, are provided in Appendix 8.3.1. (...) Our TVLG policies consist of 100 time steps and we control our robot at a frequency of 20 Hz. (...) In all of the real robot experiments, policies are updated every 10 rollouts and the final policy is obtained after 20-25 iterations, which corresponds to mastering the skill with less than one hour of experience.