Scholarlink

With the rise of embodied intelligence, leveraging large-scale human data for high-level imitation learning on humanoid robots has gained attention. However, precision operations remain challenging due to complex perception and control, fundamental human–robot morphological and actuation gaps, and the lack of task-relevant features from egocentric vision. To address covariate shift, this paper proposes an imitation learning algorithm for humanoids that: focuses on primary task objectives, filters background information, and integrates channel feature fusion with spatial attention to suppress environmental disturbances; it also employs a dynamic weight update strategy to significantly improve task success rates. Experiments show robustness and scalability across diverse tasks, offering new avenues for autonomous learning and control in humanoid robots.

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

OminiAdapt: Learning Cross-Task Invariance for Robust and Environment-Aware Robotic Manipulation

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yongxu Wang,
  • Weiyun Yi,
  • Xinhao Kong,
  • Wanting Li

摘要

With the rise of embodied intelligence, leveraging large-scale human data for high-level imitation learning on humanoid robots has gained attention. However, precision operations remain challenging due to complex perception and control, fundamental human–robot morphological and actuation gaps, and the lack of task-relevant features from egocentric vision. To address covariate shift, this paper proposes an imitation learning algorithm for humanoids that: focuses on primary task objectives, filters background information, and integrates channel feature fusion with spatial attention to suppress environmental disturbances; it also employs a dynamic weight update strategy to significantly improve task success rates. Experiments show robustness and scalability across diverse tasks, offering new avenues for autonomous learning and control in humanoid robots.