PACER: Progress-Aligned Curation for Error-Resilient Imitation Learning

Shreyas Kumar, Ravi Prakash

Keywords: Imitation Learning, Imperfect Demonstrations, Learning

TL;DR: PACER learns a latent task phase to align demos and robustly filter local corruptions, refining labels for stable imitation from few imperfect demos.

Abstract: Imitation learning from human demonstrations offers a powerful framework for robotic skill acquisition, eliminating the need for explicit reward specification. However, in realistic low-data settings, where only a handful of demonstrations are available for each task, demonstrations are not only expensive to collect but can also be imperfect: operators may experience fatigue, vary in execution strategies or timing, and occasionally introduce brief corrections or deviations, such as unintended motions or hesitations, that can occur at arbitrary stages of task execution. To address these challenges, we introduce PACER, a progress-aligned framework that aligns demonstrations in latent task phase and robustly filters local corruptions before policy learning. PACER enables reliable imitation from sparse and noisy data, yielding policies that better capture intended behavior and outperform standard behavioral cloning and alignment baselines across manipulation and locomotion domains respectively by roughly 45\% and improves episodic returns by over 50\% on average. Code available at: anonymous.4open.science/r/PACER.

Submission Number: 39