GUILD guided data collection improves imitation learning sample efficiency
Original title: GUILD: Guided Iterative Data Collection for Imitation Learning
GUILD alternates vision-language model task proposals with predictive data attribution evaluation to close the collection loop.
The GUILD framework actively reasons about data collection to maximize policy performance in imitation learning. It alternates task proposal using semantic priors from vision-language models with trajectory evaluation via predictive data attribution. This prevents uninformative data acquisition unlike post-collection filtering. Real robot and simulation evaluations show improved sample efficiency and deployment performance over baselines including experienced roboticists.
- Alternates task proposal with semantic priors from vision-language models
- Approximates policy influence via predictive data attribution
- Iterative loop closes data collection process
- Substantially improves performance versus passive collection and roboticists
From the original article
1 min read at arxiv.orgImitation learning policies in robotics are trained on datasets collected through costly real-world teleoperation. Often, the data-collection process is unguided, relying on the teleoperator's intuition, resulting in data that contributes little to overall policy performance.
By Shivin Dass · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.