An arXiv paper describes a method that chooses which teleoperated trajectories to collect, rather than filtering a finished dataset.
An arXiv paper describes a method that chooses which teleoperated trajectories to collect, rather than filtering a finished dataset.
A paper on arXiv, cs.LG, introduces GUILD, a Guided Data Collection method for imitation learning in robotics. The authors state that policies are typically trained on datasets gathered through costly real-world teleoperation, and that the process is often unguided, relying on the teleoperator's intuition, which produces data that contributes little to overall policy performance. GUILD instead actively reasons about which data should be collected to maximize downstream policy performance.
The method uses an iterative framework that alternates between task proposal with data collection and data evaluation to close the loop. For task proposal, the authors leverage semantic priors from vision-language models to identify relevant environments, object configurations, and goals for data collection. The excerpt does not report benchmark results, success rates, or dataset sizes, so no quantitative evidence is available here.
The paper notes that prior work curates unguided datasets by filtering after collection, which does not prevent acquisition of uninformative data; GUILD is presented as addressing that gap. Because semantic relevance does not imply policy improvement, the authors perform data evaluation using predictive data attribution to approximate how collected trajectories influence the policy, though the excerpt cuts off before describing that mechanism in full. The excerpt contains no experimental numbers, so readers should await the full paper before judging whether guided collection outperforms unguided collection in practice.
Source: GUILD: Guided Iterative Data Collection for Imitation Learning