One After Another: Learning Incremental Skills for a Changing World

Authors: Nur Muhammad Mahi Shafiullah, Lerrel Pinto

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

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate experimentally that in both evolving and static environments, incremental skills significantly outperform current state-of-the-art skill discovery methods on both skill quality and the ability to solve downstream tasks.
Researcher Affiliation Academia Nur Muhammad (Mahi) Shafiullah mahi@cs.nyu.edu New York University Lerrel Pinto lerrel@cs.nyu.edu New York University
Pseudocode Yes H DISK PSEUDO CODE Algorithm 2 Pseudocode for DISk training routine for the Mth skill.
Open Source Code Yes Videos for learned skills and code are made public on: https://notmahi.github.io/disk.
Open Datasets Yes for stationary tasks we use standard Mu Jo Co environments from Open AI Gym (Todorov et al., 2012; Brockman et al., 2016): Half Cheetah, Hopper, Ant, and Swimmer
Dataset Splits No No explicit percentages or counts for training/validation/test dataset splits are provided. The paper describes data collection into replay buffers for training, but not a fixed split for dataset partitioning like in supervised learning.
Hardware Specification Yes All of our experiments were run between a local machine with an AMD Threadripper 3990X CPU and two NVIDIA RTX 3080 GPUs running Ubuntu 20.04, and a cluster with Intel Xeon CPUs and NVIDIA RTX 8000 GPUs, on a Ubuntu 18.04 virtual image.
Software Dependencies Yes We base our implementation on the Py Torch implementation of SAC by Yarats & Kostrikov (2020).
Experiment Setup Yes D IMPLEMENTATION DETAILS AND HYPERPARAMETERS