Rethinking Efficiency Bottleneck of Generative Robotics Manipulation

Talk overview

This talk examines two major barriers to deploying generative embodied AI on real robots: redundant multi-view sensory data and the complexity of large embodied models. It presents compression methods for 3D/4D Gaussian Splatting and device-conditioned once-for-all model compression to reduce storage, communication, training, and inference costs.