GUILD: Guided Iterative Data Collection for Imitation Learning
GUILD: Guided Iterative Data Collection for Imitation Learning
This paper introduces GUILD, a novel method for guided data collection in imitation learning. It addresses the inefficiency of unguided teleoperation by proposing an iterative framework that alternates between task proposal and data evaluation. The system uses vision-language models to identify relevant environments and goals, then employs predictive data attribution to estimate how collected trajectories influence policy performance. This feedback loop guides subsequent data collection to maximize downstream performance. Evaluations in both simulation and real robot settings show that GUILD substantially improves sample efficiency and deployment performance compared to passive strategies, multiple baselines, and experienced roboticists.
The authors propose GUILD, an iterative framework combining semantic priors from vision-language models and predictive data attribution to actively guide teleoperation for imitation learning, demonstrating improved sample efficiency and deployment performance in both simulation and real-world settings compared to passive collection strategies.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
Source：arXiv Robotics — research abstracts · arxiv.org