DisCo RL: Distribution-Conditioned Reinforcement Learning for General-Purpose Policies

被引:3
|
作者
Nasiriany, Soroush [1 ]
Pong, Vitchyr H. [1 ]
Nair, Ashvin [1 ]
Khazatsky, Alexander [1 ]
Berseth, Glen [1 ]
Levine, Sergey [1 ]
机构
[1] Univ Calif Berkeley, Berkeley, CA 94720 USA
关键词
D O I
10.1109/ICRA48506.2021.9561402
中图分类号
TP [自动化技术、计算机技术];
学科分类号
0812 ;
摘要
Can we use reinforcement learning to learn general-purpose policies that can perform a wide range of different tasks, resulting in flexible and reusable skills? Contextual policies provide this capability in principle, but the representation of the context determines the degree of generalization and expressivity. Categorical contexts preclude generalization to entirely new tasks. Goal-conditioned policies may enable some generalization, but cannot capture all tasks that might be desired. In this paper, we propose goal distributions as a general and broadly applicable task representation suitable for contextual policies. Goal distributions are general in the sense that they can represent any state-based reward function when equipped with an appropriate distribution class, while the particular choice of distribution class allows us to trade off expressivity and learnability. We develop an off-policy algorithm called distribution-conditioned reinforcement learning (DisCo RL) to efficiently learn these policies. We evaluate DisCo RL on a variety of robot manipulation tasks and find that it significantly outperforms prior methods on tasks that require generalization to new goal distributions.
引用
收藏
页码:6635 / 6641
页数:7
相关论文
共 50 条
  • [1] Planning and Reinforcement Learning for General-Purpose Service Robots
    Jiang, Yuqian
    [J]. PROCEEDINGS OF THE THIRTIETH INTERNATIONAL JOINT CONFERENCE ON ARTIFICIAL INTELLIGENCE, IJCAI 2021, 2021, : 4895 - 4896
  • [2] Driving with Style: Inverse Reinforcement Learning in General-Purpose Planning for Automated Driving
    Rosbach, Sascha
    James, Vinit
    Grossjohann, Simon
    Homoceanu, Silviu
    Roth, Stefan
    [J]. 2019 IEEE/RSJ INTERNATIONAL CONFERENCE ON INTELLIGENT ROBOTS AND SYSTEMS (IROS), 2019, : 2658 - 2665
  • [3] Bioinspired framework for general-purpose learning
    de Toledo, SA
    Barreiro, JM
    [J]. FOUNDATIONS AND TOOLS FOR NEURAL MODELING, PROCEEDINGS, VOL I, 1999, 1606 : 507 - 516
  • [4] LEARNING ON VLSI - A GENERAL-PURPOSE DIGITAL NEUROCHIP
    DURANTON, M
    SIRAT, JA
    [J]. PHILIPS JOURNAL OF RESEARCH, 1990, 45 (01) : 1 - 17
  • [5] CONTRASTIVE LEARNING OF GENERAL-PURPOSE AUDIO REPRESENTATIONS
    Saeed, Aaqib
    Grangier, David
    Zeghidour, Neil
    [J]. 2021 IEEE INTERNATIONAL CONFERENCE ON ACOUSTICS, SPEECH AND SIGNAL PROCESSING (ICASSP 2021), 2021, : 3875 - 3879
  • [6] Reinforcement of general-purpose grade rubbers by silica generated in situ
    Kohjiya, S
    Ikeda, Y
    [J]. RUBBER CHEMISTRY AND TECHNOLOGY, 2000, 73 (03): : 534 - 550
  • [7] Machine Learning a General-Purpose Interatomic Potential for Silicon
    Bartok, Albert P.
    Kermode, James
    Bernstein, Noam
    Csanyi, Gabor
    [J]. PHYSICAL REVIEW X, 2018, 8 (04):
  • [8] Unsupervised Learning of General-Purpose Embeddings for Code Changes
    Pravilov, Mikhail
    Bogomolov, Egor
    Golubev, Yaroslav
    Bryksin, Timofey
    [J]. MALTESQUE '21: PROCEEDINGS OF THE 5TH INTERNATIONAL WORKSHOP ON MACHINE LEARNING TECHNIQUES FOR SOFTWARE QUALITY EVOLUTION, 2021, : 7 - 12
  • [9] Towards general-purpose representation learning of polygonal geometries
    Gengchen Mai
    Chiyu Jiang
    Weiwei Sun
    Rui Zhu
    Yao Xuan
    Ling Cai
    Krzysztof Janowicz
    Stefano Ermon
    Ni Lao
    [J]. GeoInformatica, 2023, 27 : 289 - 340
  • [10] Towards general-purpose representation learning of polygonal geometries
    Mai, Gengchen
    Jiang, Chiyu
    Sun, Weiwei
    Zhu, Rui
    Xuan, Yao
    Cai, Ling
    Janowicz, Krzysztof
    Ermon, Stefano
    Lao, Ni
    [J]. GEOINFORMATICA, 2023, 27 (02) : 289 - 340