Scholarlink

In the Internet of Robotic Things (IoRT), ensuring the generalizability and reliability of agents controlling robotic nodes across diverse tasks is challenging. This paper introduces the Reinforcement Learning Agent of Enhanced Generalizability (RLAEG), a novel framework for autonomous control in robotic manipulation tasks within IoRT. To enhance the agent generalizability across multiple tasks, RLAEG employs a diffusion policy, which is informed by explicit task cues provided through a dedicated multi-task auxiliary cueing module, for the generation of diverse actions. Furthermore, RLAEG introduces an action distribution training module with dynamic output guided by weighted importance sampling to further enhance agent generalizability and improve control reliability. On the CompoSuite benchmark, RLAEG achieves a significant 62.3% relative improvement in the benchmark average task success rate over the CP-IQL baseline across both uniform and restricted sampling settings. Real robot experiments with a robotic arm further demonstrate RLAEG’s potential for generalizability in autonomous robotic control tasks within IoRT environments.

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

RLAEG: A Reinforcement Learning Agent of Enhanced Generalizability for Autonomous Control in IoRT

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yukun Qian,
  • Haitao Wang,
  • Hejun Wu,
  • Zhiyang Mai,
  • Guowei Zou

摘要

In the Internet of Robotic Things (IoRT), ensuring the generalizability and reliability of agents controlling robotic nodes across diverse tasks is challenging. This paper introduces the Reinforcement Learning Agent of Enhanced Generalizability (RLAEG), a novel framework for autonomous control in robotic manipulation tasks within IoRT. To enhance the agent generalizability across multiple tasks, RLAEG employs a diffusion policy, which is informed by explicit task cues provided through a dedicated multi-task auxiliary cueing module, for the generation of diverse actions. Furthermore, RLAEG introduces an action distribution training module with dynamic output guided by weighted importance sampling to further enhance agent generalizability and improve control reliability. On the CompoSuite benchmark, RLAEG achieves a significant 62.3% relative improvement in the benchmark average task success rate over the CP-IQL baseline across both uniform and restricted sampling settings. Real robot experiments with a robotic arm further demonstrate RLAEG’s potential for generalizability in autonomous robotic control tasks within IoRT environments.