Hierarchical Policy Learning via Spectral Decomposition

2026-06-28

AI Summary

This work addresses the challenge of disentangling task intent from execution details in robotic fine manipulation by proposing Causal Spectrum Policy (CSP), which uniquely integrates semantic decomposition of action sequences with spectral characteristics. Leveraging the Discrete Cosine Transform (DCT), CSP separates low-frequency components—representing global trajectories—from high-frequency components encoding fine temporal dynamics and contact behaviors in the frequency domain, thereby establishing a causal coarse-to-fine hierarchical policy generation mechanism. Inspired by human teleoperation, the method further incorporates a noise injection strategy to enhance robustness. Experimental results demonstrate that CSP significantly outperforms strong baselines in both simulation and real-world settings, excelling particularly in precision-sensitive tasks and exhibiting strong adaptability to noisy demonstration data.

Abstract

In this paper, we identify a semantic decomposition in robot action sequences, separating task-level motion intent from execution-level refinements. By analyzing actions in the spectral domain using the discrete cosine transform (DCT), we observe that low-frequency components capture global motion trajectories, while high-frequency components encode precise timing, alignment, and contact behaviors. Motivated by this structure, we propose Causal Spectral Policy (CSP), which models action generation as a causal coarse-to-fine process: coarse motion is predicted from observation and language, and fine corrections are generated conditionally on the realized trajectory. Across simulation and real-world evaluations, CSP consistently outperforms strong baselines on precision-sensitive manipulation tasks. Additionally, we propose human-inspired teleoperation noise injection as a data augmentation method, under which our approach demonstrates strong robustness to noisy demonstrations.

Problem

• hierarchical policy learning
• spectral decomposition
• robot action sequences
• motion intent
• execution refinement

Innovation

• Spectral Decomposition
• Hierarchical Policy Learning
• Discrete Cosine Transform
• Causal Coarse-to-Fine Generation
• Teleoperation Noise Injection