A1350
Title: Physical robot-object interaction recognition using wavelet scattering-based features for transformer models
Authors: Haeun Ryu - Chungnam National University (Korea, South) [presenting]
Yunchae Jung - Chungnam National University (Korea, South)
Jiwon Im - Chungnam National University (Korea, South)
Minsu Park - Chungnam National University (Korea, South)
Abstract: Effective input representation design is essential for sensor-based action classification due to complex time-frequency structures and inherent stochastic noise in robotic signals. Conventional transformer models relying on raw signal patch-wise tokenization often exhibit vulnerability to high-frequency noise and fail to capture multi-scale temporal dependencies. To address these challenges, a robust classification framework is proposed that utilizes the wavelet scattering transform (WST) as the input representation for transformer encoders. WST provides a contraction mapping that is Lipschitz continuous to deformations, generating energy-preserving and shift-invariant features through multi-scale wavelet filtering and nonlinear operations. These mathematical properties facilitate the identification of distinct physical interaction patterns, including precision placement, kinetic release, and continuous surface contact. Evaluations using the DROID (dataset for robot instruction and demonstration) dataset demonstrate that WST-based representations achieve superior classification accuracy and enhanced noise robustness compared to traditional raw patch-based methods. The results indicate that integrating WST with transformer architectures offers a statistically sound feature engineering strategy, effectively supporting automated action classification and post-hoc process validation in autonomous robotic systems within industrial environments.
