Spectral-Temporal Conditioning: A Frequency-Domain Framework for Robust Robotic Imitation Learning

Imitation learning via trajectory generation facilitates efficient robotic skill acquisition. However, existing methods, such as Action Chunking with Transformers (ACT), suffer from cyclical failures in high-precision tasks. These failures arise because state aliasing during low-velocity phases renders consecutive observations near-indistinguishable under Markov Decision Process assumptions, and entangled frequency components in expert demonstrations drive the policy into stagnation or oscillation. To address these challenges, this paper proposes Temp-ACT, a frequency-domain framework that conditions ACT on historical trajectories via the Short-Time Fourier Transform (STFT). The framework decouples task intent from noise across three stages: an asymmetric sinusoidal window mitigates spectral leakage while preserving recent observations; a Conditional Variational AutoEncoder (CVAE) models the frequency-domain representation; and time-domain reconstruction with learnable low-pass filtering ensures output stability. Our approach enhances trajectory smoothness and progression, resolving inference stagnation in precision manipulation. Experiments in both simulation and real-world precision manufacturing validate Temp-ACT, showing that it eliminates cyclical failures and increases full-task success rates over the ACT baseline from 3% to 36% in simulation and from 5% to 30% in real-world settings.