Scholarlink

In recent decades, learning from demonstrations (LfD), a technique that develops strategies from example states to action mappings (Argall et al. in Rob Autonom syst 57(5):469–483, 2009 [1]), has attracted considerable attention along with the development of robotics and AI technologies. A recent survey on LfD concluded that current limitations of LfD include representation of complex behaviours, reliance on labelled data, and suboptimal and inappropriate demonstrators (Ravichandar et al. in Ann Rev Control Rob Autonom Syst 3:297–330, 2020 [2]). To solve this problem, this section proposes an incremental skill learning and generalization framework to enable robots to modify simple initial actions to complex cases. This method is based on motion primitive (MP) technology, in which a long-term complex motion is divided into multiple sub-actions. Then, the sub-actions are extracted into MPs and finally these MPs are reprogrammed and generalised to fit a new task (Huang et al. in Int J Rob Res 38(7):833–852, 2019 [3]).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid Learning and Control Using Improved Dynamical Movement Primitive and Adaptive Neural Network Control

  • 章节

  • 发表日期：

    2025年

查看全文
  • Chenguang Yang,
  • Zhenyu Lu,
  • Ning Wang

摘要

In recent decades, learning from demonstrations (LfD), a technique that develops strategies from example states to action mappings (Argall et al. in Rob Autonom syst 57(5):469–483, 2009 [1]), has attracted considerable attention along with the development of robotics and AI technologies. A recent survey on LfD concluded that current limitations of LfD include representation of complex behaviours, reliance on labelled data, and suboptimal and inappropriate demonstrators (Ravichandar et al. in Ann Rev Control Rob Autonom Syst 3:297–330, 2020 [2]). To solve this problem, this section proposes an incremental skill learning and generalization framework to enable robots to modify simple initial actions to complex cases. This method is based on motion primitive (MP) technology, in which a long-term complex motion is divided into multiple sub-actions. Then, the sub-actions are extracted into MPs and finally these MPs are reprogrammed and generalised to fit a new task (Huang et al. in Int J Rob Res 38(7):833–852, 2019 [3]).