FEWT: Improving Humanoid Robot Perception with Frequency-Enhanced Wavelet-based Transformers

Abstract

The embodied intelligence bridges the physical world and information space. As its typical physical embodiment, humanoid robots have shown great promise through robot learning algorithms in recent years. In this study, a hardware platform, including humanoid robot and exoskeleton-style teleoperation cabin, was developed to realize intuitive remote manipulation and efficient collection of anthropomorphic action data. To improve the perception representation of humanoid robot, an imitation learning framework, termed Frequency-Enhanced Wavelet-based Transformer (FEWT), was proposed, which consists of two primary modules: Frequency-Enhanced Efficient Multi-Scale Attention (FE-EMA) and Time-Series Discrete Wavelet Transform (TS-DWT). By combining multi-scale wavelet decomposition with the residual network, FE-EMA can dynamically fuse features from both cross-spatial and frequency-domain. This fusion is able to capture feature information across various scales effectively, thereby enhancing model robustness. Experimental performance demonstrates that FEWT improves the success rate of the state-of-the-art algorithm (Action Chunking with Transformers, ACT baseline) by up to 30% in simulation and by 6-12% in real-world.