Title: SpeedAug: Accelerating Policies Through Tempo-Enriched Prior and RL Fine-Tuning

Abstract:

Recent advancements in robotic policy learning for intricate real-world manipulation have been largely driven by the capacity to gather demonstrations via human operation. Nevertheless, policies derived from such data frequently operate at speeds significantly below the robot's physical potential. This discrepancy arises because demonstration collection is often bound by practical constraints that prioritize conservative, success-focused trajectories over rapid execution. Current methods for accelerating policies typically rely on data preprocessing or heuristic rules to set execution tempo, rather than learning a speed profile optimized for the specific task.

To address this limitation, we introduce SpeedAug, a framework that allows policies to learn task-optimal execution tempos through reinforcement learning (RL). The approach begins by training a tempo-enriched prior policy on speed-augmented demonstrations, thereby capturing a wide variety of execution speeds. Subsequently, RL fine-tuning leverages this prior to guide exploration, efficiently refining action trajectories and optimizing execution tempo. Evaluations on robotic manipulation benchmarks reveal that SpeedAug significantly enhances the sample efficiency of policy acceleration while preserving high success rates, resulting in fast and stable task performance. In a real-world application, SpeedAug increased task throughput by 1.8x, requiring only 16 minutes of online interaction and maintaining the original success rate.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC