SCIZOR: A Self-Supervised Approach to Data Curation for Large-Scale Imitation Learning (2505.22626v1)

Abstract: Imitation learning advances robot capabilities by enabling the acquisition of diverse behaviors from human demonstrations. However, large-scale datasets used for policy training often introduce substantial variability in quality, which can negatively impact performance. As a result, automatically curating datasets by filtering low-quality samples to improve quality becomes essential. Existing robotic curation approaches rely on costly manual annotations and perform curation at a coarse granularity, such as the dataset or trajectory level, failing to account for the quality of individual state-action pairs. To address this, we introduce SCIZOR, a self-supervised data curation framework that filters out low-quality state-action pairs to improve the performance of imitation learning policies. SCIZOR targets two complementary sources of low-quality data: suboptimal data, which hinders learning with undesirable actions, and redundant data, which dilutes training with repetitive patterns. SCIZOR leverages a self-supervised task progress predictor for suboptimal data to remove samples lacking task progression, and a deduplication module operating on joint state-action representation for samples with redundant patterns. Empirically, we show that SCIZOR enables imitation learning policies to achieve higher performance with less data, yielding an average improvement of 15.4% across multiple benchmarks. More information is available at: https://ut-austin-rpl.github.io/SCIZOR/

Scizor: Enhancing Data Quality in Large-Scale Imitation Learning

The paper introduces Scizor, a novel self-supervised framework aimed at improving the quality of datasets used in large-scale imitation learning. Imitation learning, an integral aspect of robotic automation, is highly dependent on the quality of data drawn from demonstration datasets which are often diverse and large-scale, leading to variability in data quality. Scizor addresses these issues by implementing a fine-grained data curation process that removes suboptimal and redundant state-action pairs, ultimately enhancing policy performance in imitation learning applications.

The crux of Scizor lies in its ability to identify low-quality data without the need for costly manual annotations. It leverages two complementary curation strategies: a suboptimal data removal module based on task progress prediction, and a deduplication module targeting redundant data patterns. The task progress predictor acts as a metric to estimate the quality of state-action sequences by determining whether the actions contribute meaningful progression towards task goals, while the deduplication module utilizes a joint state-action representation to locate and filter repetitive patterns that dilute the training signal.

Empirical evaluations demonstrate significant improvements in policy performance with Scizor's curation compared to uncurated large-scale datasets and existing curation approaches. Across diverse benchmarks in both simulated and real-world settings, Scizor achieved an average improvement in policy success rate of 15.4%, affirming its efficacy in enhancing data quality and optimizing learning outcomes. This figure indicates that Scizor effectively curates datasets, removing misleading suboptimal actions and repetitively redundant data that could otherwise skew the learning process.

The implications of this research extend to both theoretical and practical domains within AI and robotics. Theoretically, Scizor contributes to the understanding of data curation dynamics in imitation learning, emphasizing the importance of nuanced approaches to data filtering beyond traditional trajectory or dataset-level methods. Practically, the application of Scizor can be instrumental in making robotic systems more robust and reliable, ensuring they learn from optimal data and behave as intended.

Looking ahead, the potential for Scizor to adaptively calibrate its curation thresholds could further refine data quality optimization, extending its applicability across even broader datasets and tasks. Additionally, exploring learned representations for deduplication could offer deeper insights and enhancements in data curation strategies.

In conclusion, Scizor represents a significant advancement in addressing data quality issues within imitation learning, offering a scalable and effective solution for optimizing robot learning through self-supervised data curation. This approach presents promising avenues for future research and development in refining autonomous systems.