Bootstrapping Imitation Learning for Robotic Manipulation via Hierarchical Data Exploration Space

Authors:
  • JYJinrong Yang
    Jinrong Yang
  • ZLZhuoling Li
    Zhuoling Li
  • SWShengkai Wu
    Shengkai Wu
  • LRLiangliang Ren
    Liangliang Ren
  • WLWenqiu Luo
    Wenqiu Luo
  • MZMeiyu Zhi
    Meiyu Zhi
  • MSMingshan Sun
    Mingshan Sun
  • ZCZhaojie Chu
    Zhaojie Chu
  • FYFangping Yang
    Fangping Yang
  • YZYong Zhao
    Yong Zhao

Abstract
Imitation learning (IL) is a powerful paradigm for complex, long-horizon robot manipulation, yet standard IL suffers from covariate shift: small, compounding errors during autonomous execution drift the agent away from the expert's demonstration distribution, leading to task failure. This brittleness stems from data collection that relies on monolithic, repetitive demonstrations, producing datasets that are highly redundant and intolerant to minor perturbations. We propose Hierarchical Data Exploration Space (Hide-Space), a data collection framework that builds an inherently robust training distribution to preemptively mitigate covariate shift. Hide-Space decomposes long-horizon tasks into atomic sub-tasks and collects data for each by actively exploring a continuous, overlapping state space that funnels towards a goal. Evaluations on four real-world long-horizon manipulation tasks show that IL policies trained on Hide-Space data achieve significantly enhanced performance with far fewer demonstrations, and offer insights into optimizing data quality and guiding data scaling.

This publication has 4 references indexed in Scilit: