PWM: Policy Learning with Multi-Task World Models

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Georgiev, Ignat, Giridhar, Varun, Hansen, Nicklas, Garg, Animesh
Format: Preprint
Veröffentlicht: 2024
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866913703929053184
author Georgiev, Ignat
Giridhar, Varun
Hansen, Nicklas
Garg, Animesh
author_facet Georgiev, Ignat
Giridhar, Varun
Hansen, Nicklas
Garg, Animesh
contents Reinforcement Learning (RL) has made significant strides in complex tasks but struggles in multi-task settings with different embodiments. World model methods offer scalability by learning a simulation of the environment but often rely on inefficient gradient-free optimization methods for policy extraction. In contrast, gradient-based methods exhibit lower variance but fail to handle discontinuities. Our work reveals that well-regularized world models can generate smoother optimization landscapes than the actual dynamics, facilitating more effective first-order optimization. We introduce Policy learning with multi-task World Models (PWM), a novel model-based RL algorithm for continuous control. Initially, the world model is pre-trained on offline data, and then policies are extracted from it using first-order optimization in less than 10 minutes per task. PWM effectively solves tasks with up to 152 action dimensions and outperforms methods that use ground-truth dynamics. Additionally, PWM scales to an 80-task setting, achieving up to 27% higher rewards than existing baselines without relying on costly online planning. Visualizations and code are available at https://www.imgeorgiev.com/pwm/.
format Preprint
id arxiv_https___arxiv_org_abs_2407_02466
institution arXiv
publishDate 2024
record_format arxiv
spellingShingle PWM: Policy Learning with Multi-Task World Models
Georgiev, Ignat
Giridhar, Varun
Hansen, Nicklas
Garg, Animesh
Machine Learning
Artificial Intelligence
Robotics
Reinforcement Learning (RL) has made significant strides in complex tasks but struggles in multi-task settings with different embodiments. World model methods offer scalability by learning a simulation of the environment but often rely on inefficient gradient-free optimization methods for policy extraction. In contrast, gradient-based methods exhibit lower variance but fail to handle discontinuities. Our work reveals that well-regularized world models can generate smoother optimization landscapes than the actual dynamics, facilitating more effective first-order optimization. We introduce Policy learning with multi-task World Models (PWM), a novel model-based RL algorithm for continuous control. Initially, the world model is pre-trained on offline data, and then policies are extracted from it using first-order optimization in less than 10 minutes per task. PWM effectively solves tasks with up to 152 action dimensions and outperforms methods that use ground-truth dynamics. Additionally, PWM scales to an 80-task setting, achieving up to 27% higher rewards than existing baselines without relying on costly online planning. Visualizations and code are available at https://www.imgeorgiev.com/pwm/.
title PWM: Policy Learning with Multi-Task World Models
topic Machine Learning
Artificial Intelligence
Robotics
url https://arxiv.org/abs/2407.02466